Humanoid-PushRecovery-CrossEmbodiment

Physics-informed (Linear Inverted Pendulum + Capture Point) humanoid push-recovery dataset: 32 distinct robot embodiments, 1,920 trials, ankle/hip/step/fall taxonomy, non-monotonic gait-phase-dependent outcomes, and intermittent double-push disturbances. Companion any-time multi-task LSTM predicts falls and lead time from partial post-push trajectories, evaluated on unseen embodiments.

About Dataset

Humanoid-PushRecovery-CrossEmbodiment simulates balance-recovery trials for humanoid robots subjected to external pushes, terrain irregularities, and low-friction (slip) events, using the same reduced-order dynamics the push-recovery control literature itself relies on: Linear Inverted Pendulum CoM dynamics (Kajita et al., 2001), Capture Point balance-recoverability analysis (Pratt, Carff, Drakunov & Goswami, 2006), Divergent-Component-of-Motion feedback control (Englsberger, Ott & Albu-Schaffer, 2015), and the ankle/hip/step/fall strategy-escalation taxonomy of Stephens (2007). 32 independently-sampled robot "embodiments" (mass, CoM height, foot size, hip strategy budget, max step distance, sensorimotor reaction latency, sensor noise) each generate 60 trials varying push magnitude/direction, standing vs. walking gait phase, terrain roughness, and floor friction, at 50 Hz for 3 seconds (150 steps) per trial. Robot parameters are illustrative ranges spanning published research-humanoid scales, not a fitted digital twin of any named robot. Following Kalyanakrishnan & Goswami (2011, International Journal of Humanoid Robotics), the dataset reproduces their reported non-monotonic fall pattern: identical push magnitude at different points in the gait cycle produces opposite outcomes, due to the narrower base of support during single-leg stance. This is empirically verified in the generated trials. Following Mungai, Prabhakaran & Grizzle (2024, ICRA), roughly 18% of trials are "intermittent": a second, smaller push arrives while the robot is still mid-recovery from the first, testing robustness to compounding disturbances. Each timestep records 22 channels: CoM kinematics, capture point, zero-moment point, IMU-like acceleration/gyroscope readings, ankle/hip torque, ground reaction forces, simplified joint angles, and gait-support phase. The train/val/test split is grouped by robot embodiment, so entire robot designs are held out for testing generalization to unseen humanoid morphologies, not just unseen pushes on familiar hardware. This is a synthetic dataset: no real robot telemetry, proprietary, or personal data was used.

Purpose of Dataset

This Dataset Supports Research On Humanoid Fall Prediction And Balance-recovery Control That Generalizes Across Different Robot Designs, Rather Than Being Tuned To One Specific Platform. It Targets Researchers And Engineers Developing Early-warning Fall-detection Systems, Reactive Balance Controllers, And Safety Systems For Humanoid Robots, Particularly Where The Deployed Hardware May Differ From Whatever Robot Was Used During Model Development. The Grouped-by-embodiment Evaluation Design Directly Answers A Deployment-relevant Question: Does A Fall-predictor Trained On Some Humanoid Designs Transfer To A New Design It Has Never Seen A Single Trial From. This Matters As The Humanoid Robotics Industry Diversifies Across Many Distinct Hardware Platforms, Where A Generalizable Safety Model Is More Valuable Than One Tied To A Single Robot. It Also Enables Research Into Early-warning Prediction (Estimating A Fall Before It Fully Happens, From Partial Sensor Data) Rather Than Only After-the-fact Fall Detection, Which Is The More Operationally Useful Capability For Triggering Protective Responses In Real Time.