RecoverFormer: End-to-End Contact-Aware Recovery for Humanoid Robots
New end-to-end policy helps humanoid robots recover from 300N pushes and transfer zero-shot to new environments.
RecoverFormer is a new end-to-end humanoid recovery policy designed to help robots like the Unitree G1 recover from unexpected disturbances in unstructured environments. The architecture combines a causal transformer over a 50-step observation history with two novel heads: a latent recovery mode that enables smooth transitions among distinct recovery strategies (like compensatory stepping and hand-environment contact), and a contact affordance head that predicts which environmental surfaces (walls, railings, table edges) are beneficial for stabilization. This allows the robot to switch between behaviors without explicit mode supervision, validated by t-SNE analysis of 300 episodes.
Trained only on open floors, RecoverFormer transfers zero-shot to walled environments, achieving 100% recovery success across 100-300 N pushes and wall distances from 0.25 to 1.4 meters. Under zero-shot dynamics mismatch, it reaches 75.5% recovery with +25% mass, 89% under 30 ms latency, 91.5% at low friction, and 99% under compound friction, latency, and mass perturbation. These results show that a single end-to-end policy can deliver multi-modal, contact-aware humanoid recovery that generalizes across perturbation magnitude, contact geometry, and dynamics shift.
- RecoverFormer uses a causal transformer with two heads: latent recovery mode for smooth strategy transitions and contact affordance head for surface stabilization.
- Zero-shot transfer from open floors to walled environments yields 100% recovery on 100-300N pushes.
- Under compound dynamics shifts (friction, latency, mass), recovery success reaches 99%.
Why It Matters
This enables humanoid robots to recover from falls in real-world scenarios, crucial for deployment in factories and homes.