Galatea Domain — Features and Capabilities

Named after Galatea (Γαλάτεια), the ivory statue sculpted by the legendary Cypriot sculptor Pygmalion in Ovid's Metamorphoses — a creation so perfect and beloved that Aphrodite answered Pygmalion's prayer and brought it to life — Galatea is the comprehensive robotics software stack for humanoid robots operating in fashion retail and entertainment environments. Just as the mythological Galatea crossed the boundary between inanimate form and living presence, the Galatea domain gives physical robots the software intelligence to perceive, move, interact, and perform.

Galatea covers the full robotics software stack from bare-metal firmware and motor drivers through computational kinematics, balance control, computer vision, AI behavioral intelligence, ISO 13482-compliant safety systems, garment and fashion management, choreographed multi-robot show production, and fleet management — everything needed to deploy humanoid robots in commercial fashion retail and entertainment contexts.

The domain is organized as 20 module directories under libs/galatea/, building to 36 packages. It has no apps/ or services/ projects: all functionality is exported as libraries that integrations, SDKs, and future services consume.

At a Glance#

The table below summarizes the domain's 12 functional areas and which libraries provide them. Each row represents a cohesive concern; the sections that follow go into depth on each one.

Key Specifications

The table below provides a quick reference for specific technical capabilities. Detailed explanations appear in the numbered sections below.

1. Core Robotics Framework (@galatea/core)#

The foundational layer providing shared types, constants, utilities, and error definitions used by every other Galatea library. All robot configurations, coordinate frames, joint representations, and error codes originate here.

• Coordinate frame types — Standardized 3D coordinate frame definitions (base frame, body frame, end-effector frame, world frame) for consistent spatial reasoning across all robot components. Frame transformations use homogeneous matrices to represent both rotation and translation in a single operation.
• Joint angle representations — Unified joint angle data structures supporting revolute joints (rotate around an axis, like an elbow), prismatic joints (translate along an axis, like a linear actuator), and continuous joints (revolute without angle limits, like a wheel). Position, velocity, and effort fields are standardized across all joint types.
• Robot configuration schemas — Zod-validated schemas defining robot morphology (which joints exist and how they connect), joint limits (minimum and maximum angles), and operational parameters (maximum speeds, payload capacities). Configurations are validated at load time to prevent runtime failures from misconfiguration.
• Error code taxonomy — Comprehensive error code taxonomy covering kinematics failures (singularity, workspace violation, convergence failure), hardware faults (motor overtemperature, encoder fault, communication timeout), safety violations (force limit exceeded, prohibited zone entered), and communication errors.
• Configuration management — Centralized configuration system with schema validation, factory defaults, runtime override support, and configuration versioning.
• Physical constants and safety thresholds — Robot-specific physical constants (link lengths, mass properties, inertia tensors) and safety thresholds (maximum joint torques, maximum end-effector contact forces, thermal limits) exported as validated constants.
• Math utilities — Rotation matrix utilities, quaternion operations, Denavit-Hartenberg (DH) parameter calculations, and homogeneous transformation matrix composition.

2. Kinematics and Dynamics (@galatea/kinematics)#

Computational kinematics transforms between joint space (a set of joint angles) and task space (the position and orientation of the end effector in 3D space). This transformation is fundamental to all motion planning and control.

2.1 Forward Kinematics#

• Forward kinematics (FK) — Computes the end effector's position and orientation in the world frame from a given set of joint angles. Uses Denavit-Hartenberg parameter chain evaluation: a standardized method of describing kinematic chains by four parameters per joint (d, θ, a, α) that enables systematic FK computation. Essential for knowing where the robot's hand is given its current pose.
• Workspace analysis — Computes the reachable workspace boundary for any kinematic chain configuration: the set of all positions the end effector can reach with any valid joint configuration. Used to verify that planned motions are physically achievable before attempting them.

2.2 Inverse Kinematics#

Inverse kinematics (IK) solves the harder inverse problem: given a desired task target, what joint angles achieve it?

• Hierarchical whole-body IK solver — A single solver (solveWholeBodyInverseKinematics) resolves a prioritised set of tasks over the full 52+ DOF body. Each task is one of four types — end_effector (link position and/or orientation), center_of_mass, gaze (link forward axis), or posture (preferred joint angles). Tasks carry a numeric priority and optional weight and tolerance.
• Damped least-squares with null-space projection — Each priority level is solved by damped least squares (the damping term keeps joint-velocity steps bounded near singular configurations); lower-priority tasks are projected into the null space of higher-priority tasks so secondary objectives are satisfied only insofar as they do not disturb primary ones.
• Constraint enforcement — The solver clamps each iteration against joint limits, per-joint maximum velocity, and (optionally) hand-to-hand self-collision avoidance, and reports per-task convergence (satisfied, degraded, errorNorm) and the set of degraded task IDs.
• Redundancy resolution — Because the humanoid body has far more joints than any single end-effector task requires, the null-space mechanism above exploits that redundancy for secondary objectives (preferred posture, gaze) without affecting higher-priority targets.

2.3 Dynamics and Forces#

Dynamics computes the forces and torques required to produce desired motions — essential for power consumption estimation, motor sizing, and torque-based control.

• Recursive Newton-Euler dynamics — Computes joint torques required to execute any desired trajectory using the recursive Newton-Euler algorithm. This efficient algorithm computes the dynamics of an n-DOF robot in O(n) time by propagating forces and velocities outward from the base and inward from the end effector.
• Inertia tensor computation — Calculates link inertia tensors (mass distribution tensors that determine how each link resists rotation) for dynamic motion planning.
• Gravity compensation — Computes the joint torques required to hold any pose against gravity without moving. Essential for compliant gravity compensation mode where the robot resists gravity without following a stiff trajectory.
• Torque calculation — Determines the torques needed at each joint to achieve desired accelerations, accounting for gravity, Coriolis forces (arising from the interaction of rotational and translational motion), and centrifugal forces.

2.4 Collision Geometry#

• Swept volume collision detection — Detects potential collisions along planned trajectories by computing the volume swept by robot links during motion. Checks this swept volume against the environment model before execution.
• Self-collision checking — Prevents robot self-intersection during motion planning by checking proximity between all link pairs.
• Safety margin computation — Calculates minimum clearance distances between robot links and obstacles in the environment, ensuring the robot maintains safe standoff distances.
• Collision-free path planning — Plans trajectories that avoid obstacles while satisfying joint limits and trajectory smoothness constraints.

2.5 Supporting Computations#

• Jacobian computation — Computes both geometric Jacobians (relating joint velocities to end-effector linear/angular velocities) and analytical Jacobians (using Euler angle representations) for velocity-level control and singularity analysis.
• URDF/XACRO parser — Parses standard robot description formats (Unified Robot Description Format and XACRO macro files) to construct kinematic chain models. Enables the same control code to work with different robot platforms described by their URDF files.
• Singularity detection — Identifies kinematic singularities — configurations where the robot loses the ability to move in certain directions — through condition number monitoring of the Jacobian matrix. Applies damping to prevent dangerous velocity amplification near singularities.

3. Hardware Abstraction (@galatea/hardware-abstraction)#

A unified interface layer that abstracts the physical hardware details so the same control software can run on different robot platforms.

3.1 Actuator and Joint Control#

• Standardized joint interface — Unified command and feedback API across DC motors, servo motors, BLDC (brushless DC) motors, and series elastic actuators. Control software sends position, velocity, or torque commands through a common interface regardless of the underlying actuator technology.
• Actuator profiles — Characterization data for each actuator type including torque-speed curves, maximum continuous and peak torques, thermal limits, gear ratio, backlash, and encoder resolution.
• Position, velocity, and torque control modes — All three primary control modes supported on any actuator through the abstraction layer. Position mode for precise joint angle tracking; velocity mode for smooth motion at desired speeds; torque mode for compliant interaction with the environment.

3.2 Sensor Fusion#

• Multi-sensor fusion — Fuses Inertial Measurement Units (IMUs), joint encoders, force/torque sensors, and vision data into a coherent robot state estimate.
• Real-time state estimation — Combines proprioceptive sensors (joint encoders, IMUs — measuring the robot's own state) and exteroceptive sensors (cameras, depth sensors — measuring the environment) for accurate full-state estimation at control loop rates.
• Kalman and complementary filtering — Noise reduction for raw sensor data streams. Kalman filters optimally combine noisy measurements from multiple sensors; complementary filters efficiently fuse high-frequency gyroscope data with lower-frequency accelerometer data for IMU attitude estimation.

3.3 Body Systems#

• Body morphing — Parameterized body dimension configuration for different robot platforms. The same control software adapts to robots with different link lengths and segment proportions by adjusting kinematic parameters.
• Facial expression system — Controls actuators for eyes (gaze direction, pupil dilation), eyebrows, and mouth to produce recognizable emotional expressions for human-robot interaction.
• Dexterous hand control — Finger-level articulation for manipulation tasks (grasping garments, presenting objects) and gesture expression (pointing, waving).
• Tactile skin sensing — Processes capacitive tactile sensor array data distributed across the robot body for contact detection, contact force mapping, and texture discrimination.
• RFID reader interface — Reads RFID tags embedded in garments for automated outfit identification and inventory tracking without manual barcode scanning.

3.4 Thermal Management#

• Motor temperature monitoring — Real-time thermal monitoring of all motor actuators. Thermal sensors embedded in motor housings report temperatures at the control loop rate.
• Thermal throttling — Automatic performance reduction when actuators approach thermal limits: reducing maximum torque and speed to prevent overheating while allowing the robot to continue operating at reduced capability rather than shutting down.
• Electronics cooling management — Monitors and controls electronics cooling subsystems (fans, liquid cooling loops where applicable) to maintain processor and power electronics within operating temperature ranges.

4. Firmware Management (@galatea/firmware)#

Manages the embedded firmware running directly on robot microcontrollers and hardware subsystems.

4.1 Motor Drivers#

• Motor driver interface — Low-level motor control firmware handling PWM generation, current sensing, commutation sequences (for BLDC motors), and fault detection.
• Current and velocity loops — High-bandwidth inner control loops running at the microcontroller level. The current loop controls motor torque; the velocity loop controls motor speed. These run at 10–20 kHz, much faster than the higher-level kinematics loop.
• Fault detection and protection — Hardware-level protection against overcurrent, overvoltage, undervoltage, and overtemperature conditions with safe shutdown procedures.

4.2 Sensor Interfaces#

• Sensor driver library — Drivers for all supported sensors: IMUs (ICM-42688, BMI088), joint encoders (AMT22 absolute, AS5048 magnetic), force/torque sensors (ATI, Rokubi), cameras (Intel RealSense, ZED, OAK-D), and tactile arrays.
• Sensor calibration — Calibration procedures and parameter storage for each sensor type, ensuring accurate measurements across operating conditions.
• Sensor fusion firmware — Microcontroller-level sensor fusion for combining IMU accelerometer and gyroscope data at full sensor bandwidth before transmission to the main processor.

4.3 Safety Controller#

• Hardware safety controller — Dedicated safety microcontroller monitoring all safety-critical conditions independently of the main control processor. Hardware-level independence ensures safety functions work even if the main processor fails.
• E-stop handling — Processes emergency stop signals from both hardware buttons and software commands, executing safe state transitions within guaranteed time bounds.
• Watchdog timers — Hardware watchdog timers that trigger safe shutdown if the main processor stops responding within the heartbeat period.
• Safe state definitions — Defined safe states for different fault conditions (full E-stop, reduced speed mode, limp mode) with documented transition procedures.

4.4 Power Management#

• Battery state estimation — State-of-charge and state-of-health estimation for lithium battery packs using coulomb counting and voltage-based models.
• Power distribution management — Manages power distribution to subsystems with priority-based load shedding when battery is low.
• Charging interface — Handles autonomous docking and charging initiation when battery state-of-charge falls below configured thresholds.

4.5 Communication Bus Drivers#

• EtherCAT driver — Industrial real-time fieldbus driver for high-speed, synchronized communication between the main controller and motor driver subsystems. EtherCAT achieves sub-microsecond synchronization across distributed nodes.
• CAN bus driver — Controller Area Network driver for sensor data and lower-bandwidth subsystem communication.
• Serial drivers — UART and SPI drivers for direct sensor connections.

4.6 RTOS Runtime#

• Real-time operating system runtime — Bare-metal RTOS runtime for time-critical control loops. Provides task scheduling, priority management, and inter-task communication primitives with deterministic timing.
• Control loop scheduling — Configurable multi-rate control loop scheduling: motor current loops at 20 kHz, velocity loops at 1 kHz, kinematics at 500 Hz, behavior at 50 Hz.

4.7 OTA Firmware Management#

• Firmware version tracking — Tracks firmware versions across all hardware subsystems in the fleet.
• OTA firmware updates — Delivers firmware updates over the air with staged rollout (canary → regional → full fleet), rollback capability on failure, and update validation.
• Firmware compatibility checking — Verifies firmware-software compatibility before deployment to prevent mismatched version combinations that could cause unexpected behavior.
• Boot sequence management — Coordinates safe startup sequences across all subsystems with proper initialization ordering and fault handling during boot.

5. Bipedal Locomotion (@galatea/locomotion)#

Walking on two legs in unstructured environments is one of robotics' hardest problems. Galatea's locomotion system handles gait planning, real-time balance control, and terrain adaptation.

5.1 Gait Planning#

• Walking gait generation — Generates stable walking gaits at configurable speeds and step lengths. Produces joint trajectories over a complete gait cycle (stance phase + swing phase) that can be tracked by the whole-body controller.
• Running gait generation — Dynamic running patterns with aerial phases (both feet off the ground simultaneously). Requires predictive control because balance cannot be maintained at each instant.
• Configurable walking styles — Parameterizes walking style to express robot personality and match fashion context: energy-efficient industrial walk, elegant runway walk, casual natural walk, and confident presentation walk.
• Terrain-adaptive gaits — Adjusts gait parameters based on sensed floor type: carpet (higher friction, different energy recovery), ramp (incline/decline compensation), and uneven surfaces (increased foot clearance).

5.2 Balance and Stability#

• ZMP balance control — Zero Moment Point-based stability control. The ZMP is the point on the ground where the net ground reaction force effectively acts. If the ZMP lies within the support polygon (the convex hull of all ground contact points), the robot is dynamically stable. The controller continuously adjusts motion to keep the ZMP inside the support polygon.
• Center of pressure tracking — Real-time monitoring of the Center of Pressure (CoP) measured by foot pressure sensors, used as an estimator of the actual ZMP.
• Push recovery — Recovers from external disturbances (a person bumping the robot, an unexpected payload) through a hierarchy of strategies: ankle strategy (ankle torque adjustment for small perturbations), hip strategy (whole-body reconfiguration for medium perturbations), and stepping strategy (taking a step to recover for large perturbations).
• Dynamic balance — Maintains stability during dynamic tasks: reaching beyond the static stability margin, turning at speed, and carrying objects that shift the center of mass.

5.3 Navigation#

• Footstep planning — Plans optimal footstep sequences considering obstacle avoidance, gait continuity, and energy efficiency.
• Stair navigation — Ascends and descends stairs with foot placement adaptation to measured step dimensions and stair geometry.
• Obstacle avoidance — Dynamically avoids obstacles detected by the perception system during locomotion without stopping.
• Path following — Follows pre-planned global paths with smooth, dynamically consistent trajectory tracking.

5.4 Walking Styles#

A purpose-built library of walking style parameterizations for fashion and retail contexts:

• Runway walk — High-energy, deliberate, theatrical walk appropriate for fashion show presentation. Exaggerated stride, elevated step height, confident pace.
• Natural walk — Biomechanically natural walking style for approachable retail interaction.
• Quiet locomotion — Noise-minimized locomotion policy reducing mechanical noise for quiet retail environments where intrusive sound would be inappropriate.

6. Whole-Body Motion Control (@galatea/whole-body-control)#

Whole-body motion control coordinates all joints simultaneously to achieve complex tasks that require the entire body to work together — reaching far while maintaining balance, gesturing while walking, or presenting a garment while tracking a customer.

• Task-space control — Controls end effectors (hands, head, gaze) in Cartesian space while the underlying joint motion is automatically coordinated. Motion commands are specified in natural task-space terms (move hand to this position) rather than individual joint angles.
• Impedance control — Variable impedance control where the robot behaves like a spring-damper system with configurable stiffness and damping. Low impedance makes the robot compliant and safe for human contact; high impedance provides precise position control for structured manipulation.
• Admittance control — Force-to-motion control: external forces applied to the robot cause it to move in proportion to the force. Makes the robot feel yielding and safe when people touch it — critical for human-robot interaction in public retail spaces.
• Postural control — Maintains desired body posture during task execution using null-space optimization. The robot can reach forward with its arm while simultaneously maintaining an upright, aesthetically pleasing torso posture.
• Center of mass control — Tracks and regulates the whole-body Center of Mass position and velocity for stability during dynamic whole-body movements.
• Angular momentum control — Regulates angular momentum during dynamic motions — turning, reaching, and recovery — preventing the robot from spinning out of control during fast maneuvers.

6.1 Task-Space Controller Sub-System#

• Priority-based task composition — Multiple simultaneous tasks (balance, reach, gaze) executed with a strict priority hierarchy: safety always overrides balance; balance always overrides reaching; reaching overrides gaze.
• Null-space exploitation — Uses kinematic redundancy to satisfy lower-priority tasks in the null space of higher-priority tasks, achieving all objectives simultaneously when possible.
• Task switching — Smooth transitions between task sets without discontinuous joint accelerations.

6.2 Impedance and Admittance Control#

• Variable stiffness profiles — Predefined stiffness profiles for different interaction contexts: rigid for precision manipulation, compliant for human contact zones, stiff for load-bearing.
• Contact force estimation — Estimates contact forces from joint torque measurements when no dedicated force/torque sensors are present.

7. Pose Engine (@galatea/pose-engine)#

The pose engine provides a rich library of named, semantically meaningful poses and the tools to blend between them smoothly — enabling the robot to express meaning through body language.

7.1 Pose Library#

• Named pose library — Catalog of semantically named poses for fashion and retail contexts: attention (alert, ready to assist), welcome (open, inviting), present_garment (highlighting the outfit being worn), bow (formal greeting), wave (casual greeting), thinking (considering), and many more.
• Biomechanically valid poses — All library poses are validated against joint limits and self-collision constraints. Invalid poses cannot be added to the library.
• Motion capture retargeting — Imports pose data from motion capture recordings (BVH, FBX, C3D formats) and retargets them to the Galatea skeleton, allowing fashion professionals and choreographers to author poses using familiar tools.

7.2 Pose Interpolation and Transitions#

• Pose interpolation — Smooth blending between two poses with configurable duration, easing curve (linear, ease-in, ease-out, cubic bezier), and interpolation path.
• Pose sequencing — Chains multiple poses with timing, easing, and hold durations into complete expressive motion sequences.
• Transition planner — Plans dynamically feasible transitions between poses, respecting joint limits and velocity constraints throughout the motion.
• Context-aware pose selection — Selects contextually appropriate poses based on the current task (presenting garments vs. greeting a customer vs. standing on the runway), interaction state, and sensed audience engagement level.

7.3 Natural Motion Generation#

• Breathing simulator — Adds continuous subtle breathing motion (slow rhythmic torso expansion and chest rise) to any static pose, making the robot appear alive rather than frozen.
• Micro-movement generator — Adds subtle continuous micro-movements (weight shifts, small head adjustments, natural body sway) that make the robot feel natural and present rather than mechanically static between commanded motions.
• Contrapposto solver — Computes the classical contrapposto pose (one hip raised, opposite shoulder raised, natural S-curve through the spine) — the foundational aesthetic pose of Western figurative art and fashion. Enables the robot to stand in this naturally appealing stance rather than mechanical symmetry.
• Pose optimizer — Optimizes poses for visual aesthetics and physical stability simultaneously, finding the best achievable pose given both constraints.

8. Computer Vision and Perception (@galatea/perception)#

Galatea's perception system enables robots to understand the people, environment, and garments around them.

8.1 People and Audience Awareness#

• Person detection — Detects people in the robot's camera field of view with 2D bounding boxes and 3D position estimation using depth data.
• Person tracking — Tracks individual people over time as they move through the retail space, maintaining consistent IDs across occlusion events.
• Audience awareness — Detects audience presence, group size, and spatial distribution during fashion shows and retail demonstrations.
• Engagement estimation — Estimates audience engagement levels from gaze direction, body orientation toward the robot, proximity, and dwell time.

8.2 Garment and Fashion Perception#

• Garment recognition — Classifies garment type (dress, jacket, trousers, etc.), color, pattern, and style from camera images.
• Fit analysis — Analyzes garment fit on the robot using visual measurement, detecting whether garments are correctly positioned and seated.
• Fashion trend analysis — AI-powered analysis of fashion trends and style compatibility from visual observations in the retail environment.

8.3 Environment and Navigation Perception#

• SLAM (Simultaneous Localization and Mapping) — Builds a map of the retail environment while simultaneously localizing the robot within it. Enables the robot to navigate without pre-installed markers or GPS.
• Obstacle detection — Detects both static obstacles (display cases, walls, fixtures) and dynamic obstacles (people, shopping carts) for safe navigation.
• Depth processing — Processes Intel RealSense and ZED stereo depth camera data for 3D scene understanding, including floor plane detection and obstacle height estimation.
• Visual servoing — Image-based visual servo control that uses camera feedback to precisely position the end effector relative to a visual target — useful for garment presentation alignment and interaction targeting.
• Attention prediction — Predicts where the robot should direct its gaze and attention next based on scene understanding and social cues.

9. AI and Robot Intelligence (@galatea/ai)#

9.1 Foundation Models for Robotics#

Foundation models pre-trained on large robot datasets provide general robotic capability that can be fine-tuned for specific tasks.

• Vision-Language-Action (VLA) runtime — Executes VLA models that map visual observations and natural language instructions directly to motor actions. VLA models enable robots to follow natural language commands ("pick up the blue jacket") without programming each action explicitly.
• VLA fine-tuning pipeline — Fine-tunes VLA models on Galatea-specific data (garment manipulation, fashion presentation, customer interaction) to improve task performance beyond the general pre-trained baseline.
• Large Behavior Model (LBM) — Foundation behavior model fine-tuning for complex, long-horizon tasks requiring multi-step reasoning: "Greet the customer, present the outfit, describe the key features, and offer to show an alternative."
• LBM evaluation suite — Systematic evaluation of behavior model performance across standardized task scenarios, enabling quantitative comparison of model versions.
• Motor cortex policy — Hierarchical control policies for coordinated multi-joint motion, implementing learned motor primitives (reach, grasp, wave, bow) that can be composed by higher-level behavior systems.

9.2 Behavioral Systems#

• Behavioral engine — Finite state machine and behavior tree execution engine for structured robot behavior. Behavior trees enable hierarchical composition of behaviors with clear priority, fallback, and sequencing semantics.
• Natural motion generation — Generates natural-looking motions from high-level behavioral descriptions ("wave hello enthusiastically", "present this garment elegantly") without requiring explicit trajectory programming.
• Emotion expression — Generates emotional expressions through facial features, body posture, and movement timing to communicate robot states (welcoming, attentive, delighted, apologetic) that support natural human-robot interaction.
• Customer engagement — Multi-turn customer interaction system with dialog management, context tracking, and natural conversation capability for retail assistance scenarios.
• LLM integration — Large language model integration for conversational ability, knowledge-based responses (describing garment materials, care instructions, outfit styling advice), and natural language instruction following.

9.3 Learning and Training Infrastructure#

• Reinforcement learning infrastructure — RL training for locomotion, manipulation, and interaction policies. Supports both on-robot learning and simulation-to-real transfer.
• End-to-end neural control — Neural network control policies mapping directly from sensor input to motor output, trained through RL or imitation learning.
• Data collection — Systematic robot experience data collection during deployment, capturing state-action-outcome tuples for offline policy improvement.
• Distributed training infrastructure — Large-scale distributed training for robot policies requiring significant compute.
• Simulation farm management — Manages fleets of simulation instances running in parallel for efficient RL policy training.

9.4 Remote Operation#

• Teleoperation — Remote control with real-time video streaming and optional force feedback for the operator. Used for demonstration, data collection, and operational rescue.
• VR teleoperation — VR headset-based teleoperation with intuitive motion mapping from operator movements to robot movements. The operator's body motions directly drive the robot.
• Sensor suit integration — Captures human motion from operator sensor suits for natural teleoperation and motion capture data collection for imitation learning.

10. Safety and Compliance (@galatea/safety)#

Safety is non-negotiable for robots operating near people in public spaces. Galatea's safety systems are designed to meet ISO 13482, the international standard for the safety requirements of personal care robots.

10.1 Standards Compliance#

• ISO 13482 compliance checking — Automated compliance checking against the personal care robot safety standard (ISO 13482:2014). Systematically verifies that the robot's design and operational parameters satisfy each requirement.
• Automated risk assessment — Systematic hazard identification, risk estimation (severity × probability), and risk evaluation following ISO 12100 risk assessment methodology.
• Regulatory documentation toolkit — Generates CE (European conformity) marking documentation, UL (Underwriters Laboratories) certification documentation, and market-specific compliance packages from the risk assessment results.

10.2 Real-Time Safety Systems#

• Force and torque limiting — Real-time limits on contact forces and joint torques preventing injury during human-robot contact. Limits are configurable per body region (head: very low, arm: medium, tool: higher) following ISO/TS 15066 contact force and pressure limits.
• Emergency stop system — Hardware and software E-stop with guaranteed safe state transition times. Physical E-stop buttons on the robot override all software states.
• Functional safety monitoring — Continuous safety monitoring with watchdog timers, heartbeat checking, and safety channel redundancy.
• Safe state transitions — Defined, tested safe states entered on any safety violation detection: immediate stop (all joints hold position), controlled stop (gentle deceleration to zero velocity), and power-off (gradual power removal in safe sequence).

10.3 Operational Safety#

• Speed and separation monitoring — Monitors human proximity using depth cameras and reduces robot speed when people are within defined safety zones. Stops the robot if humans enter the minimum safety zone.
• Safety zone enforcement — Configurable safety zone geometry (spherical, cylindrical) around the robot with response policies (reduce speed, pause, stop) triggered at each zone boundary.
• Fault detection and diagnosis — Detects and diagnoses hardware and software faults before they become safety hazards. Distinguishes recoverable faults (temporary sensor noise) from non-recoverable faults requiring safe shutdown.

11. Garment and Fashion Management (@galatea/garment-management)#

Purpose-built capabilities for fashion retail operation, representing Galatea's domain-specific differentiation.

• Outfit tracking — Real-time tracking of which garments the robot is currently wearing or displaying, using RFID reads and visual confirmation.
• RFID garment identification — Reads garment RFID tags for automated outfit change logging. As the robot changes outfits, the system automatically records which garments were added or removed.
• Quick-change system — Coordinates rapid outfit transitions during fashion shows and retail demonstrations. Choreography scripts can include outfit change cues with precise timing.
• Cloth manipulation — Control algorithms for handling soft, deformable garments: picking up a jacket by the collar, smoothing a dress's hem, presenting a scarf. Cloth manipulation is one of robotics' hardest problems due to the infinite degrees of freedom of fabric.
• Garment damage detection — Detects potential garment damage during handling (excessive tension, contact with sharp edges, inappropriate grip force) and aborts operations to prevent damage.
• Outfit inventory management — Tracks the full garment inventory available to the robot: which garments are available, their location in the garment storage area, and their condition.

12. Choreography and Show Production (@galatea/choreography)#

Multi-robot choreography for fashion shows and entertainment performances — enabling coordinated, synchronized, aesthetically polished robot performances at scale.

• Choreography definition — Defines complex multi-robot choreographies with per-robot motion sequences, timing, formations, and transition rules.
• Music synchronization — Synchronizes robot motions to music tracks with beat detection (identifying strong beats for accent movements), BPM analysis, and configurable motion-music alignment.
• Show scripting — Scripts complete show sequences with lighting cue triggers, music playback control, robot positions, and outfit change timing.
• Multi-robot coordination — Coordinates multiple robots performing simultaneously with collision avoidance and formation maintenance.
• Formation management — Defines and executes formation changes: from single-file runway walk to arc facing the audience to symmetric pairs. Manages the transition choreography between formations.
• Rehearsal mode — Executes shows at reduced speed (configurable fraction of real speed) for rehearsal and debugging without time pressure.
• Show analytics — Records show performance data (timing adherence, motion quality metrics, audience engagement signals) for quality review and choreography improvement.

13. Fleet Management (@galatea/fleet)#

Manages the operational lifecycle of multiple deployed robots across retail locations and show venues.

• Fleet registry — Registers and tracks all robots in the fleet with identity (serial number, name), capability profile (hardware version, supported tasks), and deployment history.
• Remote monitoring — Real-time telemetry from all fleet robots: joint states, battery level, temperature, current task, error log, and position.
• Fleet health dashboard — Aggregated fleet health metrics and alerts: how many robots are operational, how many need maintenance, how many are in error states.
• Remote configuration — Deploys configuration updates to individual robots or robot groups without physical access.
• Maintenance scheduling — Schedules and tracks preventive maintenance based on operating hours, actuator cycle counts, and condition-based maintenance triggers.
• Software update management — OTA software and firmware update deployment with staged rollout, validation tests, and rollback capability on failure.
• Deployment assignment — Assigns robots to retail locations, show venues, or specific roles within a venue.

14. Robot Communication (@galatea/communication)#

Communication infrastructure for robot-to-cloud and robot-to-robot interaction.

• Publish-subscribe messaging — Topic-based message bus for robot telemetry and command distribution. Robots publish sensor data and state; the cloud subscribes to relevant topics and publishes commands.
• Low-latency command channel — Real-time command delivery optimized for teleoperation and safety stops where latency directly affects safety. Uses dedicated connection to guarantee delivery time.
• Robot-to-robot coordination — Direct inter-robot communication for multi-robot show coordination. Robots share position, task state, and timing signals directly rather than relying solely on cloud coordination.
• Event streaming — Streams robot events to cloud processing pipelines for analytics, logging, and model training data collection.

15. Database and Persistence (@galatea/database)#

Five PostgreSQL-backed store packages. Each generates CREATE TABLE/INDEX DDL, provides parameterised query-template builders, and exposes a store class that runs against an injected database client.

• event-store — Append-only operational event log spanning the safety, incident, operator_action, ota_update, and operational domains, with integrity verification.
• telemetry-store — TimescaleDB hypertable store for joint-state, battery, and environmental telemetry, with 1m/1h aggregates, retention policy, and sampling-profile configuration.
• garment-store — Garment catalog, RFID mappings, digital-product-passport cache, wear history, and fit data.
• pose-store — Named-pose library supporting vector similarity search and full-text search.
• show-store — Show definitions, their version history, and show execution records.

16. Event Handlers (@galatea/event-handlers)#

Five sub-packages, each a stateful in-process handler engine with a frozen event-type constant array and a replayable, filterable event history.

• Robot lifecycle events (robot-events) — Models seven robot subsystems with boot-priority dependency graphs; runs boot sequences and self-tests, classifies and isolates faults, selects recovery strategies, escalates, and persists shutdown state. Emits 20 robot event types.
• Safety event handlers (safety-events) — Handles E-stop triggers, manual resets, force-limit-approached warnings, collision detection, and stability-margin-low events (five safety event types).
• Show event handlers (show-events) — Registers show definitions, tracks robot readiness, starts shows, dispatches lighting / formation-change / outfit-reveal cues, handles in-show robot faults, and ends shows (eight show event types).
• Garment event handlers (garment-events) — Maintains a garment catalog and handles RFID/NFC tag reads, dressed/undressed lifecycle, and digital-product-passport scans (four garment event types).
• Customer event handlers (customer-events) — Handles customer approach, engage, and depart events; produces pose and head-turn commands and manages a voice-interaction session and interaction metrics.

17. Simulation and Digital Twin (@galatea/simulation)#

Simulates robots before physical deployment and enables behavior testing without hardware risk. Eight simulation modules are unified under a single library, covering physics through show preview.

• Physics simulation — High-fidelity rigid body dynamics simulation for robot testing. Simulates joint dynamics, contact physics, and environmental interactions using an RTOS-accurate plant model for hardware-in-the-loop equivalence.
• Cloth simulator — Soft body physics simulation of clothing using finite element methods, enabling cloth manipulation algorithm testing without risking real garments or requiring physical robot time.
• Digital twin synchronization — Maintains a live virtual replica of each deployed robot with state synchronized from real-time telemetry. The digital twin visualizes the robot's current state without requiring a camera view, enabling remote operators to understand what a robot is doing at any moment.
• Show preview — Full choreography preview system that renders a complete multi-robot show in simulation with correct music synchronization and formation geometry before committing to physical rehearsal. Catches timing errors, collision risks, and formation problems at zero hardware cost.
• RL training environment — Reinforcement learning training environment with physics-accurate dynamics for training locomotion, manipulation, and interaction policies through simulated experience before deployment to real hardware.
• Wear simulator — Simulates garment wear patterns and fabric deformation under repeated outfit changes, enabling prediction of garment longevity and identification of manipulation sequences that cause premature garment damage.
• Virtual showroom — Interactive 3D virtual showroom rendering for visualizing robot–garment interactions before physical deployment, usable by fashion designers and retail buyers to evaluate how specific garments will appear on a humanoid robot.
• Scenario tester — Parameterized scenario runner for systematic testing of specific situations (customer approaches from the left while robot is presenting, two customers simultaneously request assistance) across a sweep of scenario variants.

18. Inclusivity and Accessibility (@galatea/inclusivity)#

Ensures that robotic systems serve all people equitably, regardless of body type, ability, or cultural background.

• Diverse body profiles — Body dimension profiles covering a wide range of human body types for garment fitting demonstrations and inclusive retail interaction.
• Accessibility configurations — Configures robot behavior for users with mobility differences (adjusting interaction height, slowing motion near wheelchair users), visual differences (providing audio description of what the robot is doing), or hearing differences (relying on visual communication rather than spoken language).
• Cultural configuration — Adjusts robot greetings, gestures, and interaction styles for different cultural contexts. A bow is appropriate in Japanese contexts; a handshake in Western contexts; a nod may be more appropriate in certain contexts than either.
• Multilingual support — Customer interaction in multiple languages, configured per deployment location.

19. Analytics and Business Intelligence (@galatea/analytics)#

Measures the business impact of robot deployments and enables data-driven optimization.

• Engagement analytics — Tracks customer engagement metrics per robot deployment: dwell time near the robot, interaction rate (proportion of nearby customers who interact), return visit rate, and positive vs. negative interaction sentiment.
• Show performance analytics — Measures audience reaction and engagement during choreographed performances: applause detection, attention tracking, social media mention volume, and post-show survey correlation.
• A/B testing framework — Compares robot behavior variants (different greeting styles, different show choreography, different garment presentation techniques) and measures conversion and engagement differences with statistical significance testing.
• Revenue attribution — Attributes sales lift to robot interactions and demonstrations using experimental design (control days without robots vs. robot deployment days).
• Operational efficiency metrics — Tracks robot uptime, task completion rate, error rate, maintenance costs, and total cost of deployment.

20. Developer SDK (@galatea/sdk)#

Tools for building applications and extensions on the Galatea platform.

• TypeScript SDK (@galatea/sdk) — GalateaClient composes three sub-clients — fleet, shows, and analytics — over a pluggable GalateaClientBackend. The default InMemoryGalateaClientBackend wires the client to the fleet orchestrator, show designer, show scheduler, and the engagement / heatmap / revenue analytics engines, so applications can register robots, author and schedule shows, and query analytics through one typed surface.
• Python client (galatea-client-sdk) — Python counterpart of the TypeScript SDK (fleet, shows, analytics clients over an in-memory backend) plus an additional MLOpsClient for ML workflows.
• Show SDK (@galatea/sdk/show-sdk) — ShowAuthoringSdk provides a fluent ShowAuthoringBuilder for show drafts (path/pose/pivot tracks, formations, music-sync points, lighting cues, outfit-change triggers), a validateDraft step that checks track requirements, robot proximity, and the one-microsecond timing-synchronisation target, and deployDraft for validated deployment.
• Analytics SDK (@galatea/sdk/analytics-sdk) — Custom-event schemas, aggregate/ratio metric definitions, metric evaluation, and dashboard widget creation (KPI, time-series, leaderboard) for custom reporting.
• Fleet event subscription — GalateaClient.fleet.subscribeRobotStateUpdates polls the orchestrator event stream and pushes FleetRobotStatusUpdates to a listener, so applications can react to robot status changes without managing the event cursor themselves.

21. Planned Features#

The Phase 63 library surface is in place. The following capabilities are planned as part of future Galatea development and are not yet implemented:

• Advanced VLA fine-tuning (planned) — Robot-specific fine-tuned VLA models with larger training datasets from deployed robot operation.
• Cross-domain integration with Themis (planned) — Autonomous-systems governance frameworks for robotic operation in public spaces, including safety governance, incident reporting, and regulatory compliance documentation.
• Extended fashion show formats (planned) — Interactive show formats where robots respond to audience input in real time, enabled by enhanced audience engagement sensing and LLM-based natural interaction.
• Expanded retail environment coverage (planned) — Pre-built simulation environments for additional retail formats.

Planned Cross-Domain Integrations#

Galatea currently has no cross-domain code dependencies — no @galatea/* package imports another Oshun domain. The boundaries below represent future integration points rather than current coupling.

The boundary exists intentionally: Galatea is a self-contained robotics platform. The partner domains (Neith, Gaia, Nous, Themis) own their respective primitives and will expose them through well-defined APIs. Galatea will consume those APIs without becoming coupled to their internal implementations.

• Neith digital-human and grooming (Phases 155, 162) (planned) — Consume Neith hair/grooming primitives (curve hair, grooming brushes, simulation, shading, presets) and digital-human primitives (character creation, FACS facial rig, body rig/deformation, soft-tissue simulation, digital garments, performance capture, retargeting, consent/likeness/provenance) for fashion-show visualisation, robot digital twins, choreography preview, and inclusive body-profile modelling.
• Gaia weather safety inputs (Phase 175) (planned) — Consume Gaia outdoor-weather and cyclone products as safety inputs for Galatea fleets.
• Nous world models and continual learning (Phases 176-177) (planned) — Consume Nous robotic world models for sample-efficient real-robot fine-tuning, imagination safety shields, and sim-to-real adaptation, plus continual-learning and interpretability gates for fleet-wide skill retention.