Terrain Adaptation

This is the robot's internal sense of its body state. By fusing data from inertial measurement units (IMUs), joint encoders, and force/torque sensors, the system estimates critical values in real-time:

• Base pose and velocity (orientation, angular rate, linear acceleration)
• Joint positions and velocities
• Contact states (foot strike, lift-off, slippage)

This fused state is the foundational input for all subsequent planning and control loops, allowing the robot to know if it is tilting, slipping, or bearing unexpected loads.

This is the robot's external perception of the world. Using depth cameras, LiDAR, or stereo vision, the system builds a local map of the upcoming terrain. Key processed features include:

• Surface normals and slope estimation
• Step height and negative obstacles (holes)
• Terrain classification (rigid, deformable, slippery)
• Foothold scoring based on geometry and predicted stability

This forward-looking perception enables predictive adaptation, allowing the robot to plan foot placements several steps ahead rather than reacting only at the moment of contact.

The robot dynamically adjusts the spatial and temporal parameters of its cyclic walking pattern. This is not re-planning the entire gait but online optimization of:

• Stride length and step frequency: Shortening strides on slippery ground, increasing frequency for rapid traversal.
• Swing foot clearance: Lifting the foot higher to clear obstacles.
• Duty factor: Adjusting the ratio of stance phase to swing phase for stability on rough terrain.
• Body attitude: Tilting the torso to keep the center of mass over the support polygon on slopes.

These modulations allow the same underlying gait generator (e.g., a trot) to be effective across a wide range of surfaces.

Upon foot contact, the robot must manage interaction forces. This layer provides the final, millisecond-level adjustment.

• Impedance Control: Adjusts the virtual spring-damper behavior of each leg. On soft ground, it increases compliance (softer spring) to prevent digging in; on hard ground, it increases stiffness for precise positioning.
• Force Control: Regulates the ground reaction force (GRF) profile. On ice, it limits horizontal shear forces to prevent slippage; on uneven ground, it distributes weight to maintain balance.
• Reflexes: Triggers pre-programmed, fast reactions to events like unexpected foot slip (rapid leg retraction) or sudden load loss (leg search).

This component coordinates all the robot's joints to achieve locomotion objectives while respecting physical limits. Formulated as a Model Predictive Control (MPC) or Quadratic Program (QP) solved at high frequency (50-500 Hz), it optimizes for:

• Center of Mass (CoM) trajectory that maintains dynamic balance.
• Swing foot trajectories that avoid obstacles.
• Contact force distribution that satisfies friction cones.
• Joint torque limits and kinematic constraints.

It integrates the desired gait parameters, perceived terrain, and state estimate to produce dynamically consistent motion plans for the entire body.

Increasingly, the adaptation logic itself is learned. Deep Reinforcement Learning (RL) is used to train neural network control policies in high-fidelity physics simulators where the robot experiences millions of trials on randomized terrain. The learned policy encodes complex adaptation behaviors, such as:

• Recovering from severe slips and pushes.
• Navigating highly deformable substrates (e.g., sand, mud).
• Exploiting terrain features for leverage.

The core challenge is sim-to-real transfer, using techniques like domain randomization and dynamics randomization to ensure the policy's robustness transfers to the physical robot, closing the perception-action loop with learned intelligence.

This algorithm continuously replans the robot's foot placements based on real-time perception of the upcoming terrain. It uses a cost map generated from depth sensors to evaluate potential footholds. Key criteria include:

• Surface normal for slope stability
• Height variance to avoid edges
• Terrain classification (e.g., rigid vs. soft) The planner selects the foothold that minimizes a combined cost of instability, energy, and deviation from the global path. This is the primary method for navigating discrete obstacles like stepping stones.

Instead of tracking rigid position trajectories, this method controls the dynamic relationship between a leg's motion and the contact force, making the robot behave like a tunable mass-spring-damper system. The controller adjusts leg stiffness and damping in real-time based on estimated terrain properties.

• High stiffness is used on solid ground for precise positioning.
• Low stiffness (high compliance) is used on uncertain or soft terrain to absorb shocks and prevent foot slipping. This allows the robot to maintain stable contact and adapt its posture without explicit knowledge of every ground irregularity.

This advanced approach uses a dynamic model of the robot's center of mass and angular momentum (centroidal dynamics) to optimize future motions. An MPC solver runs at high frequency (50-100 Hz) to:

• Predict body motion over a short horizon (~0.5 seconds).
• Optimize ground reaction forces for each foot.
• Adjust the body's center of mass trajectory and foot forces to maintain balance on uneven terrain. It explicitly accounts for friction cones and torque limits, making it robust to pushes and slippery surfaces.

Before adapting, the robot must estimate key terrain properties. This is done by fusing pre-contact vision (e.g., a RGB-D camera) with proprioceptive sensing during contact.

• Pre-contact: A convolutional neural network classifies visual patches (e.g., 'gravel', 'grass', 'metal grate').
• During contact: The robot estimates ground stiffness and friction coefficient by observing the relationship between commanded leg motion and measured force. These estimated parameters are then fed directly into the footstep planner and impedance controller to tailor the gait.

Different gaits are optimal for different terrains. Adaptation systems can automatically trigger a gait transition.

• A static walk is used for precise, slow navigation on very rough terrain.
• A dynamic trot is used for efficient, faster travel on moderately uneven ground.
• A bound or gallop might be used to traverse soft, dissipative terrain like sand. The system modulates gait parameters within a chosen gait, such as step height, stride length, and duty factor (percentage of time a foot is on the ground), to overcome obstacles or sinkage.

Instead of hard-coded controllers, Deep Reinforcement Learning (RL) is used to train a neural network policy end-to-end for terrain adaptation. Trained in diverse simulated environments, the policy learns to map proprioception (joint angles, velocities) and exteroception (depth scans) directly to joint torques.

• Sim-to-Real Transfer: The policy is trained with domain randomization (varying friction, masses, visuals) to work on physical hardware.
• Advantage: Discovers highly robust and non-intuitive recovery behaviors.
• Challenge: Requires massive simulation compute and careful reward function design to ensure stability.

The Zero-Moment Point (ZMP) is a fundamental dynamic stability criterion for legged locomotion. It is defined as the point on the ground where the net moment of the inertial and gravitational forces acting on the robot has no horizontal component. If the ZMP remains within the support polygon (the convex hull of ground contact points), the robot is dynamically stable. This principle is central to the control of many statically stable walking robots, where terrain adaptation involves adjusting foot placements and body motion to keep the ZMP within the stable region despite ground height variations.

Reactive locomotion refers to control strategies that generate immediate, reflex-like adjustments to a robot's gait without re-planning a full trajectory. This is essential for terrain adaptation on highly dynamic or uncertain ground. Key techniques include:

• Push recovery: Using ankle, hip, or stepping strategies to counteract unexpected external forces.
• Compliant leg control: Modifying leg stiffness and damping in real-time based on estimated ground properties.
• Reflexive foot placement: Adjusting swing foot trajectory mid-step upon detecting an obstacle or hole via foot sensors. These strategies operate on a faster timescale than full motion re-planning, providing robustness to sudden terrain changes.

Whole-Body Control (WBC) is a hierarchical optimization framework that coordinates all of a robot's degrees of freedom to execute multiple tasks simultaneously while respecting physical constraints. For terrain adaptation, a WBC solver might prioritize tasks in this order:

1. Maintain contact constraints (feet cannot penetrate the ground).
2. Track desired center of mass motion for balance.
3. Achieve specific foot placements on uneven terrain.
4. ​Regulate body posture (e.g., keep torso level). It typically formulates these competing objectives as a Quadratic Program (QP), solved at high frequency (e.g., 1 kHz) to compute optimal joint torques that adapt the entire body to the terrain geometry.

Model Predictive Control (MPC) is an advanced method used for terrain-adaptive locomotion planning. At each control cycle, MPC:

• Uses a dynamic model (often a simplified Reduced-Order Model like the Linear Inverted Pendulum) to predict the robot's future state over a horizon of 1-2 seconds.
• Solves an optimization problem to find the sequence of future control actions (e.g., footstep locations, body accelerations) that minimizes a cost (e.g., energy, deviation from a path) while satisfying constraints (e.g., friction cones, joint limits).
• Executes only the first control action, then re-solves with new sensor data. This receding horizon approach allows the robot to proactively plan stable footsteps on upcoming rough terrain, rather than just reacting to it.

The Ground Reaction Force (GRF) is the force vector exerted by the ground on a robot's foot. It has a normal component (supporting weight) and frictional components (preventing slip). The Center of Pressure (CoP) is the point on the contact surface where the total GRF is considered to act. For terrain adaptation, accurate estimation and control of the GRF and CoP are critical:

• Force distribution: On uneven ground, the WBC or MPC must solve how to distribute total required forces among the feet to maintain balance.
• Slip prevention: The controller must ensure the tangential GRF stays within the friction cone defined by the terrain's coefficient of friction.
• Stability monitoring: The CoP location relative to the support polygon's edge is a direct measure of stability margin.

A Reduced-Order Model (ROM) is a simplified dynamic representation that captures the essential locomotion dynamics while ignoring the full robot's complexity. They are the "engine" inside real-time planners like MPC for terrain adaptation. Common ROMs include:

• Linear Inverted Pendulum Model (LIPM): Assumes constant center of mass height; used for walking on relatively flat but uneven ground.
• Spring-Loaded Inverted Pendulum (SLIP): Models the leg as a spring; useful for adapting to compliant or deformable terrain (e.g., sand, grass) in running/hopping robots.
• Angular Momentum Pendulum Model: Accounts for centroidal angular momentum; important for adaptation strategies that use swinging arms or legs for balance on slopes. These models are tractable for fast optimization, allowing the robot to plan hundreds of steps ahead to navigate complex terrain.