Proprioception

Joint encoders are sensors that measure the angular position and, by differentiation, the velocity of a robot's rotational joints. They are the primary source for a robot's kinematic state.

• Types: Absolute encoders provide a unique position value on power-up; incremental encoders measure relative change.
• Simulation Model: Encoders are simulated by directly reading the ground-truth joint angle from the physics engine and adding configurable quantization noise (based on bit resolution) and Gaussian white noise.
• Key Data: Joint position (θ in radians), joint velocity (ω in rad/s).
• Example: A 16-bit absolute encoder on a revolute joint has a resolution of 0.0055 degrees (360°/65536).

An IMU is a sensor package containing accelerometers and gyroscopes (and often magnetometers) that measure a body's specific force and angular rate.

• Accelerometers: Measure linear acceleration (m/s²), including gravity. In free fall, they read 0 m/s².
• Gyroscopes: Measure angular velocity (rad/s) around the sensor's axes.
• Simulation Challenge: Accurately modeling bias (a constant offset), bias instability (slowly drifting bias), scale factor errors, and wideband noise (e.g., angle random walk, velocity random walk) is critical for realistic sim-to-real transfer.
• Key Data: Raw specific force (a_x, a_y, a_z), raw angular rate (ω_x, ω_y, ω_z), and temperature (for bias modeling).

These sensors measure the internal forces within a robot's structure, crucial for force control and contact detection.

• Joint Torque Sensors: Measure the torque (Nm) being generated at or applied to a specific joint, often located in series-elastic actuators.
• Six-Axis Force-Torque (F/T) Sensors: Typically mounted at the wrist or foot, they measure three orthogonal forces (Fx, Fy, Fz) and three orthogonal torques (Tx, Ty, Tz).
• Simulation Model: Torque is computed via the physics engine using inverse dynamics or read directly from simulated joint actuators. Models include cross-axis coupling and hysteresis.
• Application: Essential for impedance control, collision detection, and measuring ground reaction forces for legged robots.

In many electric actuators, the motor current is a direct proxy for output torque, governed by the motor's torque constant (K_t).

• Principle: Torque (τ) ≈ K_t * Current (I). This allows for torque control without a dedicated joint torque sensor.
• Simulation Model: The simulated current is derived from the commanded torque or the computed physics interaction, then passed through a model of the motor driver and electrical dynamics (e.g., PWM frequency, back-EMF).
• Noise Sources: Includes electrical noise and quantization from the analog-to-digital converter (ADC).
• Key Data: Phase currents (I_a, I_b, I_c) or a single equivalent current value.

Raw sensor data is fused using state estimation algorithms to produce a reliable, high-frequency estimate of the robot's full body state.

• Core Algorithm: The Extended Kalman Filter (EKF) or its variants is the standard for legged and mobile robots. It combines:
  • IMU data (high-frequency, drifts).
  • Encoder data (low-drift, low-frequency kinematics).
  • Kinematic model (the robot's URDF/SDF).
• Output: A fused estimate of base position, base orientation (quaternion), base linear/angular velocity, and joint positions/velocities.
• Simulation Role: The filter is tested in sim with realistic noise models. Ground truth from the simulator is used to validate and tune the estimator before real-world deployment.

The accuracy of a simulated proprioceptive sensor model is validated and calibrated through system identification (SysID).

• Process: Real sensor data (e.g., IMU noise profiles, encoder quantization) is collected from physical hardware. Statistical models (e.g., Allan variance for IMUs) are fit to this data.
• Parameter Injection: These identified parameters (noise sigma, bias drift model, friction coefficients) are injected into the simulation's sensor models.
• Goal: Minimize the reality gap by ensuring the distribution of simulated sensor readings matches the distribution of real sensor readings. This is a prerequisite for successful sim-to-real transfer of learned policies.

Exteroception is a robot's perception of the external world, acquired through outward-facing sensors. It is the complementary sense to proprioception, forming a complete perceptual system.

• Key Sensors: Cameras, LiDAR, radar, microphones, and proximity sensors.
• Function: Detects objects, maps environments, recognizes events, and provides context for navigation and manipulation.
• Integration: Effective robotics requires sensor fusion to combine exteroceptive data ("where is the object?") with proprioceptive data ("where is my hand?") for coherent action.

Sensor fusion is the computational process of combining data from multiple, often disparate, sensors to produce a more accurate, complete, and reliable state estimate than any single sensor could provide.

• Core Algorithms: Techniques include the Kalman filter (for linear systems), extended and unscented Kalman filters (for non-linear systems), and particle filters.
• Application in Proprioception: Fuses data from joint encoders, IMUs, and torque sensors to resolve ambiguities (e.g., encoder drift vs. IMU noise) and estimate true joint position, velocity, and load.
• Challenge: Requires precise sensor calibration to align data into a common reference frame.

Kinematics describes motion without considering forces. It is the geometric foundation for interpreting proprioceptive data and planning movement.

• Forward Kinematics (FK): Calculates the end-effector position and orientation given all joint angles. This transforms proprioceptive joint data into a usable Cartesian workspace pose.
• Inverse Kinematics (IK): Calculates the required joint angles to achieve a desired end-effector pose. This is the critical planning step that converts a task goal into proprioceptive setpoints for the controller.
• The Jacobian: A matrix that linearly relates joint velocities to end-effector Cartesian velocity, essential for velocity control and force analysis.

Dynamics describes motion considering the forces that cause it. It is essential for accurate simulation and high-performance torque control.

• Forward Dynamics: Computes the resulting joint accelerations given applied joint torques and the current state (position, velocity). This is the core calculation in a physics simulator.
• Inverse Dynamics: Computes the joint torques required to achieve a desired joint acceleration. It is used in model-based control schemes like computed-torque control.
• Role in Simulation: Accurate actuator models and friction models are required within the dynamics equations to generate realistic proprioceptive signals (like torque feedback) for training.

These are advanced control paradigms that rely on high-fidelity proprioception, particularly torque sensing, to achieve compliant and force-sensitive interactions.

• Torque Control: A low-level mode where the controller commands a desired motor current/torque directly. This allows the robot to exert specific forces but requires precise joint torque feedback.
• Impedance Control: Regulates the dynamic relationship between end-effector position error and contact force. It makes the robot behave like a programmable mass-spring-damper system, crucial for safe human interaction and delicate assembly tasks.
• Simulation Need: Training these controllers requires simulating the full actuator dynamics and contact models to generate realistic proprioceptive feedback loops.

Calibration is the process of determining the accurate parameters of a sensor or sensor suite. Uncalibrated sensors render proprioceptive data useless for precise control.

• IMU Calibration: Determines sensor biases (constant offset errors) and scale factors for accelerometers and gyroscopes. This is often done via static and multi-position averaging.
• Kinematic Calibration: Measures the exact geometric parameters of a robot arm (link lengths, joint axes misalignments) to improve the accuracy of the kinematic model used by FK/IK.
• Simulation Context: In Sim-to-Real transfer, the statistical noise and bias models applied to simulated IMU and encoder outputs are derived from real-world calibration data of the physical sensors.