RecoverFormer: End-to-End Contact-Aware Recovery for Humanoid Robots
RecoverFormer employs causal Transformer with latent mode and contact prediction, achieving 100% success in humanoid recovery, zero-shot transfer, and robustness under unseen disturbances.
Key Findings
Methodology
This approach uses a causal Transformer encoder over 50 observation steps, integrating a latent recovery mode predictor and a contact affordance head. The encoder processes proprioceptive data, contact distances, and previous actions, capturing temporal dependencies. The latent mode, learned via entropy and utilization regularizers, segments recovery strategies without supervision. The contact head predicts the stabilization potential of environment surfaces, enabling proactive bracing. Training employs PPO with reward functions encouraging fast recovery, environmental exploitation, and robustness. Experiments demonstrate zero-shot transfer from open floor to walled environments, with high success rates under various perturbations, validating the model’s ability to implicitly identify and adapt to different force regimes and environmental affordances.
Key Results
- In MuJoCo simulations, RECOVERFORMER achieves 100% recovery success across 50–300N pushes on open floor, and maintains 100% success in wall environments at similar forces. Under unseen disturbances like +25% mass, 30ms latency, and low friction, success rates remain above 75%, reaching 99% under combined perturbations. The model learns distinct latent modes corresponding to force regimes, validated via t-SNE analysis, which cluster according to force intensity and recovery outcome. These results demonstrate strong generalization, robustness, and interpretability of the learned strategies.
Significance
This work advances humanoid balance recovery by integrating multi-modal, contact-aware decision-making within an end-to-end framework. The ability to transfer zero-shot to unseen environments and maintain high robustness under model mismatch addresses longstanding challenges in deploying humanoid robots in real-world, unstructured settings. The approach reduces reliance on offline contact databases and pre-planned strategies, enabling more autonomous, adaptable, and resilient robots. Such capabilities are crucial for applications like disaster response, service robotics, and human-robot collaboration, where unpredictable disturbances are common. The combination of Transformer-based temporal modeling and unsupervised strategy segmentation offers a new paradigm for robust robot control, bridging the gap between simulation and real-world deployment.
Technical Contribution
The paper introduces a novel causal Transformer architecture that encodes a 50-step observation history, enabling implicit system identification. It incorporates a latent recovery mode predictor learned via regularization, which segments recovery behaviors without supervision, providing interpretability. The contact affordance head predicts the stabilization potential of environmental surfaces, facilitating proactive bracing. The entire policy is trained end-to-end with PPO, integrating multiple modalities and environmental cues. This design allows the model to handle unseen dynamics, such as friction variations and latency, without online adaptation. The approach demonstrates that a single, unified policy can perform multi-modal, contact-aware recovery with high robustness and generalization, surpassing prior methods that relied on offline contact databases or fixed strategies.
Novelty
This research is the first to apply a causal Transformer for end-to-end humanoid fall recovery, integrating latent behavior segmentation and environment contact prediction. Unlike prior works that depend on offline contact databases or predefined strategies, this model learns to identify and switch between recovery modes autonomously, driven solely by reward signals. Its zero-shot transfer to environments with different contact geometries and dynamics marks a significant step forward. The unsupervised learning of meaningful recovery modes and the implicit system identification capability distinguish it from existing approaches, offering a flexible, scalable solution for complex, real-world scenarios.
Limitations
- The model is trained solely in simulation, and real-world deployment may encounter perception noise, sensor delays, and unmodeled dynamics, requiring domain adaptation techniques.
- Performance in highly cluttered or dynamic environments remains untested; current validation focuses on static walls and open floors.
- Inference relies on a 50-step history, which may impose computational overhead, limiting real-time deployment on resource-constrained hardware.
Future Work
Future efforts will focus on transferring the model to physical humanoid robots, incorporating visual perception for environment understanding, and developing online adaptation mechanisms. Enhancing robustness in highly dynamic or cluttered scenarios, integrating multi-modal sensory data, and reducing computational costs are key directions. Additionally, combining this approach with higher-level planning and reasoning modules could enable more complex behaviors like obstacle avoidance and task execution, further broadening the applicability of contact-aware, end-to-end humanoid control in real-world applications.
AI Executive Summary
Humanoid robots operating in unstructured environments face significant challenges in maintaining balance after unexpected disturbances. Traditional control methods often rely on offline planning, precomputed contact databases, or fixed strategies, limiting their adaptability and robustness. These approaches struggle to generalize across diverse scenarios, especially when environmental conditions or robot dynamics deviate from assumptions. To address this, recent research has explored end-to-end learning frameworks, but most lack the ability to reason about environmental contact opportunities or switch between multiple recovery behaviors seamlessly.
This paper introduces RECOVERFORMER, a novel contact-aware recovery policy based on a causal Transformer architecture. The core innovation lies in encoding a 50-step observation history to implicitly identify the robot’s dynamic state and environmental context. It predicts a latent recovery mode, which segments recovery strategies into interpretable categories, and a contact affordance score indicating beneficial surfaces like walls or railings. The entire system is trained using PPO with a reward function that encourages rapid recovery, environmental exploitation, and robustness to disturbances.
Experimental results in MuJoCo simulations demonstrate remarkable zero-shot transfer capabilities. Trained only on open-floor data, RECOVERFORMER achieves 100% success in recovering from pushes of 50–300N, and maintains high success rates (>75%) under unseen disturbances such as increased mass, latency, and reduced friction. Notably, it also generalizes to environments with walls at varying distances, actively bracing against surfaces when beneficial. The learned latent modes correlate strongly with force regimes, validating their interpretability.
This work significantly advances humanoid robotics by enabling a single, unified policy to perform multi-modal, contact-aware recovery in diverse scenarios. Its robustness and transferability open pathways for deploying humanoid robots in real-world, unpredictable environments, with applications spanning disaster response, service robotics, and autonomous assistance. Future directions include physical robot validation, environment perception integration, and online adaptation, pushing the frontier of autonomous humanoid resilience.
Deep Dive
Abstract
Humanoid robots operating in unstructured environments must recover from unexpected disturbances-a capability that remains challenging for end-to-end control policies. We present RECOVERFORMER, a fully end-to-end humanoid recovery policy that learns when and how to switch among recovery behaviors-including compensatory stepping, hand-environment contact, and center-of-mass reshaping-while maintaining robust performance under model mismatch. 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, and a contact affordance head that predicts which environmental surfaces (walls, railings, table edges) are beneficial for stabilization. We evaluate RECOVERFORMER on the Unitree G1 humanoid in MuJoCo. Trained only on open floor, RECOVERFORMER transfers zero shot to walled environments, achieving 100% recovery success across 100-300 N pushes and across wall distances from 0.25-1.4m. Under zero-shot dynamics mismatch, RECOVERFORMER reaches 75.5% at plus +25% mass, 89% under 30 ms latency, 91.5% at low friction, and 99% under compound friction, latency and mass perturbation. The learned latent modes specialize across force regimes without mode-level supervision, validated by t-SNE analysis of 300 episodes. Taken together, 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.