Push Recovery

The Ankle Strategy is the first line of defense against small perturbations. It treats the robot as an inverted pendulum and maintains balance by generating a torque at the ankle joint to shift the Center of Pressure (CoP) within the support polygon. This is a local, in-place adjustment that does not require stepping.

• Mechanism: The controller applies corrective ankle torques to keep the Zero-Moment Point (ZMP) inside the support polygon.
• Limitation: Only effective for small pushes where the required restoring torque does not exceed the friction limits of the foot-ground contact.

The Hip Strategy is activated for larger disturbances where ankle torque alone is insufficient. It uses rapid torso rotation (hip actuation) to generate angular momentum that counteracts the push, effectively swinging the upper body to keep the Center of Mass (CoM) over the feet.

• Mechanism: A quick flexion or extension of the hips creates a reaction force that opposes the disturbance.
• Dynamic Effect: This strategy leverages the centroidal angular momentum of the robot. It is faster than taking a step but consumes more energy than the ankle strategy.
• Example: A human leaning their torso forward after a push to the chest uses a hip strategy.

The Step Strategy is the primary recovery method for significant pushes. It involves taking one or more rapid steps to re-establish a viable support polygon under the robot's displaced Center of Mass (CoM).

• Core Concept: Based on the Capture Point theory. The robot calculates the point on the ground where placing a foot will allow it to come to a complete stop.
• Implementation: Requires fast swing leg trajectory generation and precise foot placement control. The step location is optimized to maximize stability, often using the Divergent Component of Motion (DCM) for planning.
• Variants: Includes single-step recovery, multiple-step stumble recovery, and lateral stepping.

Model Predictive Control (MPC) formulates push recovery as a real-time optimization problem. Using a dynamic model (like the Linear Inverted Pendulum Model), it predicts future states over a horizon and computes optimal ground reaction forces and footstep locations to maintain stability.

• Advantage: Proactively plans a sequence of recovery actions, considering future constraints, rather than reacting to the immediate state.
• Output: At each control cycle (e.g., 1-5 ms), MPC solves a Quadratic Program (QP) to determine the optimal immediate control action (e.g., CoP adjustment) and the next best footstep.
• Use Case: This is the dominant method in modern humanoid robots like Boston Dynamics' Atlas for handling large, complex disturbances.

Whole-Body Control (WBC) coordinates all the robot's joints to execute recovery as a multi-task optimization problem. It can simultaneously manage balance (a high-priority task) with other tasks like arm swing for momentum control or maintaining a gaze direction.

• Hierarchy: Tasks are ranked by priority. Balance is typically the highest priority task, formulated as keeping the ZMP stable.
• Capability: Can generate complex, whole-body motions that use the arms and torso for counter-balancing, which is more effective than using only the legs.
• Formulation: Often solved as a sequence of Quadratic Programs (QPs) that satisfy task objectives, dynamics equations, and physical constraints (joint limits, friction cones).

These strategies bypass explicit dynamic models in favor of ultra-fast reactions or learned policies.

• Reactive Reflexes: Pre-programmed, state-triggered actions. For example, a detected high torso angular velocity instantly triggers a predefined stepping motion. These are simple and fast but not optimal for all scenarios.
• Learning-Based Recovery: Uses Reinforcement Learning (RL) or Imitation Learning to train a neural network policy that maps sensor states (IMU, joint angles) directly to joint torques or actions. The policy learns complex recovery behaviors, including stumbling and fall prevention, through millions of trials in physics-based simulation.
• Sim-to-Real Transfer: The trained policy is transferred to the physical robot using domain randomization and robust training techniques.

The Atlas humanoid robot demonstrates highly dynamic push recovery using a combination of whole-body impulse control and model predictive control (MPC). When pushed, it executes rapid, coordinated movements of its limbs and torso to redirect angular momentum and regain balance. Its recovery strategies include:

• Angular momentum management to counteract rotational disturbances.
• Stepping reflexes that place feet outside the nominal gait pattern to create a new support polygon.
• Compliant actuation to absorb impact forces without breaking contact. This enables recovery from significant lateral and frontal pushes, showcasing the maturity of reactive balance control.

The quadruped robot ANYmal employs reinforcement learning (RL)-trained policies for robust push recovery on challenging terrain. Researchers at ETH Zurich and ANYbotics have trained neural network controllers in simulation that transfer directly to the physical robot. Key aspects include:

• End-to-end policy learning where the network maps proprioceptive state (joint angles, IMU data) directly to joint torque commands.
• Domain randomization during training to bridge the sim-to-real gap, varying dynamics parameters like friction and motor latency.
• The policy learns complex recovery maneuvers, including cross-step recovery and crouching, without explicit programming of the stepping logic. This approach generalizes to unseen disturbances.

Research at the MIT Biomimetics Lab has pioneered contact-implicit planning and reactive control for push recovery. The Cheetah 3 quadruped uses a mixed-integer convex optimization solver to plan body motions, foot contact forces, and foot placement over a future horizon in real-time (≈30 ms). For the Mini Cheetah, a simpler but faster balance controller adjusts the center of mass trajectory and foot forces using a quadratic program (QP) formulation. This allows recovery from kicks and shoves while running, emphasizing the role of high-frequency control (500-1000 Hz) and accurate state estimation.

The Florida Institute for Human and Machine Cognition (IHMC) has developed walking controllers for bipeds like NASA's Valkyrie that use the Divergent Component of Motion (DCM) and Capture Point theory for push recovery. Their method involves:

• Online adjustment of the DCM trajectory in response to a measured disturbance.
• Re-planning of footstep locations to place the next Capture Point to arrest the robot's falling motion.
• Upper body angular momentum management via a whole-body controller to keep the torso upright. This provides a strong theoretical foundation for step recovery, translating complex dynamics into an intuitive planning problem.

In industrial settings, robots like Boston Dynamics' Spot and ANYbotics' ANYmal use push recovery to maintain operational continuity during remote inspection tasks. If bumped by personnel or equipment, the robot must recover without falling to avoid damage and continue its mission. Implementations typically use:

• Hierarchical controllers where a high-level balance module overrides the nominal walking gait.
• Force/torque sensing in the legs to detect unexpected contact and estimate the disturbance.
• Pre-defined recovery behaviors (e.g., a controlled squat or lateral step) that are triggered based on the magnitude and direction of the push. This ensures functional safety and reliability in semi-structured environments like oil rigs and construction sites.

The DARPA Robotics Challenge (DRC) was a pivotal real-world test for humanoid push recovery. Teams had to operate robots in degraded conditions, where falls were costly. This spurred advances in:

• Static stability margins: Many teams used conservative, slow walking with a large support polygon to minimize fall risk.
• Active balancing: Robots like DRC-Hubo would widen their stance or use ankle strategies to compensate for small disturbances.
• Self-righting: A critical capability demonstrated was the ability to recover from a fall, which is a final-line push recovery strategy. The DRC highlighted that recovery isn't just about avoiding a fall, but also managing the aftermath, influencing subsequent research in fall mitigation and safe falling techniques.

The Zero-Moment Point (ZMP) is the point on the ground where the net moment of the inertial and gravitational forces acting on the robot has no horizontal component. It is a foundational criterion for dynamic stability in legged locomotion.

• Primary Use: Stability analysis and gait generation for quasi-static walking.
• Key Limitation: Assumes the foot remains flat on the ground (no rotation), making it less suitable for highly dynamic motions like running.
• Relation to Push Recovery: A controller often tries to keep the ZMP within the support polygon to prevent a fall. A strong push can rapidly shift the ZMP toward the polygon's edge, triggering recovery steps.

The Capture Point is a point on the ground where a legged robot can place its foot to come to a complete stop in one step, based on the Linear Inverted Pendulum Model (LIPM).

• Calculation: Derived from the robot's current Center of Mass (CoM) position and velocity.
• Core Function: Provides a target for foot placement during reactive push recovery. Stepping to the instantaneous capture point is an optimal strategy to arrest falling motion.
• Practical Application: Modern humanoid robots use real-time capture point calculations to decide where to step when pushed, rather than relying on pre-planned trajectories.

The Divergent Component of Motion (DCM) is a state variable that captures the unstable part of the Center of Mass dynamics in the Linear Inverted Pendulum Model.

• Mathematical Definition: ξ = x + (ẋ / ω), where x is CoM position, ẋ is CoM velocity, and ω is the pendulum's natural frequency.
• Key Insight: The DCM naturally diverges from the CoM over time. Controlling the DCM (often by placing the foot at the Capture Point, which is the DCM's projection on the ground) directly manages instability.
• Control Paradigm: DCM-based controllers explicitly separate stable and unstable dynamics, providing a powerful framework for planning recovery steps and stabilizing walking gaits.

Whole-Body Control (WBC) is a hierarchical control framework that coordinates all robot joints to execute multiple tasks simultaneously (e.g., balance, foot tracking, arm manipulation) while respecting physical constraints like torque limits and contact forces.

• Core Mechanism: Formulates control as a Quadratic Program (QP) that solves for optimal joint torques or accelerations every control cycle.
• Role in Push Recovery: When a push is detected, a WBC framework can dynamically re-prioritize tasks. For example, maintaining balance becomes the primary task, temporarily sacrificing precise foot trajectory tracking to generate a stabilizing step or torso motion.
• Advantage: Enforces physical consistency across all joints, ensuring the recovery motion is dynamically feasible.

Model Predictive Control (MPC) is an advanced method that uses an internal dynamic model to predict future system behavior over a finite time horizon and solves an optimization problem at each time step to determine optimal control inputs.

• Predictive Nature: Allows the robot to plan ahead for recovery, not just react. It can optimize a sequence of foot placements and body motions to absorb a disturbance.
• Application to Locomotion: CoM trajectory, footstep locations, and Ground Reaction Forces (GRFs) are common optimization variables.
• Push Recovery Use Case: An MPC controller can preemptively adjust the gait for upcoming pushes if predicted, or compute the optimal multi-step recovery strategy after a disturbance is measured, considering future state constraints.

Centroidal dynamics describes the relationship between the net external forces and moments acting on a robot and the motion of its Center of Mass (CoM) and its centroidal angular momentum.

• Key Equation: Relates the time derivative of the robot's centroidal momentum (linear and angular) to the sum of external wrenches (forces and torques).
• Critical for Dynamic Motions: Essential for analyzing and controlling jumping, running, and recovery where significant angular momentum is involved.
• Connection to Push Recovery: A sudden push applies an external wrench, causing a rapid change in centroidal momentum. Recovery strategies, especially angular momentum management (e.g., swinging arms or torso), are derived from centroidal dynamics principles to cancel this unwanted rotation.