Emergency stop (E-stop) mechanisms are the de facto standard for robot safety. However, for humanoid robots, abruptly cutting power can itself cause catastrophic failures; instead, an emergency stop must execute a predefined fallback controller that preserves balance and drives the robot toward a minimum-risk condition. This raises a critical question: from which states can a humanoid robot safely execute such a stop? In this work, we formalize emergency stopping for humanoids as a policy-dependent safe-stoppability problem and use data-driven approaches to characterize the safe-stoppable envelope. We introduce PRISM (Proactive Refinement of Importance-sampled Stoppability Monitor), a simulation-driven framework that learns a neural predictor for state-level stoppability. PRISM iteratively refines the decision boundary using importance sampling, enabling targeted exploration of rare but safety-critical states. This targeted exploration significantly improves data efficiency while reducing false-safe predictions under a fixed simulation budget. We further demonstrate sim-to-real transfer by deploying the pretrained monitor on a real humanoid platform. Results show that modeling safety as policy-dependent stoppability enables proactive safety monitoring and supports scalable certification of fail-safe behaviors for humanoid robots.
Figures & tables
Fig. 2: Data-efficient stoppability monitoring via PRISM. Exhaustively collecting failure data to learn SSE boundaries on physical humanoids is prohibitively expensive and risks severe hardware damage. To overcome this labeling bottleneck, PRISM dynamically tracks prediction errors and allocates dense sampling to highly uncertain boundary regions (the IS Window). By focusing on critical unstoppable states rather than trivially safe regions, PRISM reduces data requirements while improving boundary characterization. Red dots denote counts, while the shaded background indicates density values.
ri(k)=ℓ(V^stop(k)(xi),yi),(xi,yi)∈Dval.
Algorithm 1 PRISM: P roactive R efinement of I mportance-sampled S toppability M onitors
Fig. 3: Experimental platforms used : ( Left ) Real humanoid deployment. ( Middle ) AR based of data collection. ( Right ) Simulation-based fallback policy rollouts. AR-based data collection improves safety during real-robot experiments by projecting the virtual environment onto the real robot, enabling realistic data collection while performing collision checking in simulation to avoid physical collisions.
Fig. 4: Safety score over a full loco-manipulation trajectory. The monitor exhibits precise temporal alignment with the ground truth and identifies elevated risks during manipulation phases ( Pick , Place ) while ensuring high confidence during steady-state locomotion ( Transfer , Leave ).
Fig. 5: Spatial projection of the stoppability boundaries. For both (a) Real-World and (b) Simulation, the scatter plots (left panels) display the Cartesian (x,y) state distributions colored by ground-truth labels. The contour maps (right panels) illustrate the interpolated predictive safety regions overlaid with state density.
Deploying humanoid task policies requires reliable runtime monitors that anticipate unrecoverable falls in time to trigger a fallback. We formulate this task as learning the stochastic, finite-horizon recoverability envelope of a given fallback policy. Synthesizing this monitor is challenging: existing methods assume training data naturally covers the recoverability envelope, yet acquiring such data remains a major bottleneck for contact-rich systems. Consequently, they recover only a fraction of the true envelope, leaving them prone to optimistic hallucinations for out-of-distribution (OOD) inputs. We resolve this via a unified framework that iterates between value learning and active exploration. Data exploration is reframed as a reach-avoid problem, where an uncertainty-guided explorer systematically uncovers unseen states while avoiding confidently unrecoverable regions. This loop is further supported by a combined Temporal Difference and Monte Carlo objective for accurate in-distribution value prediction and reliable uncertainty estimation. By explicitly treating the remaining high-uncertainty regions as unrecoverable, the monitor mitigates optimistic OOD hallucinations. Evaluated on stochastic benchmarks and a 29-DoF Unitree G1 humanoid, our approach drastically improves monitor coverage and accuracy and successfully transfers zero-shot to real-world hardware.
Safe control of humanoid robots remains challenging due to their high-dimensional dynamics, contact-rich interactions, and sensitivity to disturbances. Although reinforcement learning has enabled effective locomotion and motion tracking, learned policies can still generate unsafe actions that lead to instability or falls. In this work, we propose residual reinforcement learning as an implicit safety-filtering mechanism for safe humanoid control. Instead of relying on a single nominal policy to simultaneously balance performance, safety, and robustness, we decouple performance and safety. The nominal policy focuses solely on task performance, while a residual policy learns safety corrections. This decoupling leads to a better performance--safety Pareto trade-off and avoids the need for careful tuning of multiple competing reward terms within a single policy training. We show that the residual policy can act as an implicit safety filter.
Gechen Qu, Tong Zhang, Bike Zhang +4
University of California, Berkeley · University of California, Los Angeles
Dynamic humanoid motions such as flips risk hardware damage due to suboptimal policies, disturbances or sim-to-real gaps. A motion tracking policy offers no way out once the maneuver leaves its reference, and a backup policy needs to take over to protect the hardware for a minimum-damage landing. Which backup to use matters as much as when to switch. We present Viability-Aware Policy Selection (VAPS), which treats safety as a policy-conditioned, receding-horizon decision. Besides a protective fall policy, we also train an abort policy which can abort the motion at any time, landing on its feet. At every control step, learned predictors estimate whether the nominal tracking policy and the abort policy remain viable over a short horizon, and a least-sacrificial hierarchy keeps the most task-ambitious behavior that remains viable. In simulation with randomized disturbances, VAPS sharply reduces head contact and hand contact, which are the dominant sources of hardware damage, with both a Unitree G1 and a LimX Oli; on the LimX Oli, we validate the viability predictors and the full VAPS controller for side-flip motions. VAPS Pareto-dominates the strongest single-network alternatives we could train, including an end-to-end safe-tracking policy and students distilled from VAPS's own oracle-routed decisions, in both task success and head impact. We also show that VAPS is a powerful framework to supervise undertrained policies and protect the hardware.
Siwei Ju, Lu Liu, Jan Peters +1
Department of Computer Science, Technical University of Darmstadt, Germany. · Robotics Institute Germany (RIG). · LimX Dynamics. +1