Continue, Abort, or Fall: Viability-Aware Policy Selection (VAPS) for Safe Humanoid Acrobatics
Preprint. VAPS is a safety layer for humanoid acrobatic robots that adds two backup policies—an abort policy that lands on its feet and a protective-fall policy—and selects among them plus the nominal tracking policy using short-horizon viability predictors. In simulation on a LimX Oli, it cut head-contact rate from 15.5% to 1.1% while keeping motion success nearly unchanged (79.2% vs 80.7%), and on the physical robot the full hierarchy achieved its intended objective in 15 of 18 side-flip trials. The work matters because it directly addresses the main cost of deploying dynamic humanoids: hardware damage from failed maneuvers.
How can a robot decide in real time whether to continue an acrobatic motion, abort to a controlled landing, or commit to a protective fall, so that hardware damage is minimized without unnecessarily giving up task success?
Dynamic humanoid maneuvers such as flips are fast and unforgiving; if a motion-tracking policy leaves its reference due to disturbances or sim-to-real gaps, there is little time to recover, and a single backup policy (usually a safe fall) cannot adapt to intermediate states where a controlled landing is still possible but the task is already lost. Existing binary switches between nominal control and one backup therefore either escalate too early and sacrifice successful maneuvers, or too late and allow severe falls that damage the head and hands.
Prior work typically learns reference-tracking policies for acrobatics and pairs them with a single fall-prevention or protective-fall behavior triggered by a learned fall predictor (e.g., SafeFall), a safety critic, control barrier function, reachability, or model-predictive safety filter; some recent methods fold execution and recovery into one end-to-end policy. These approaches compute one safety value for one policy or use a binary handover, so they cannot distinguish between states where aborting upright is feasible and states where only a fall remains, and monolithic policies struggle to balance conflicting objectives, often failing to trigger recovery at the right time.
VAPS decomposes safe execution into three specialist behaviors—nominal tracking, abort (feet-first upright landing), and protective fall (minimize head/hand impact)—each trained in simulation with domain randomization from switch states drawn from the nominal policy’s failures. For the nominal and abort policies, separate learned viability predictors estimate the probability of staying failure-free over a short horizon (e.g., 50 control steps) from a small history of proprioceptive observations; they are trained with a discounted cross-entropy loss over future offsets. At each control step, a least-sacrificial rule selects the most task-ambitious policy whose predicted viability exceeds a calibrated threshold (false-alarm rate ≤1%), and transitions are irreversible, so an abort that becomes non-viable escalates to protective fall. The authors also train monolithic baselines—end-to-end RL and distilled students of VAPS—to test whether the structure can be collapsed.
On the Unitree G1 in simulation under randomized disturbances, nominal-only success was 62.1% with 27.7% head contact and 37.9% fall rate; full-hierarchy VAPS kept success at 59.6% (−2.5 pts) while reducing head contact to 1.6%, hand contact to 8.8%, and fall rate to 26.1%. On LimX Oli simulation, VAPS full hierarchy reduced head contact from 15.5% to 1.1%, hand contact from 19.3% to 4.8%, and fall rate from 19.3% to 9.1%, with success 79.2% vs 80.7%. VAPS’s viability predictor achieved 97.5% episode accuracy, 91.7% fall detection, 0% false alarms, and mean lead time 0.416 s, versus 85.8% detection and 0.205–0.316 s lead time for SafeFall-style baselines. Against monolithic policies, the best end-to-end RL kept head contact at 24.0% (vs VAPS 1.6%) and the best distilled student reached 10.1% head contact but with lower success (56.3% vs 59.6%). In hardware on LimX Oli: the predictor classified all 13 side-flip trials correctly with mean lead 0.544 s; the abort policy landed upright in 17/19 manual-switch trials and the protective fall protected head/hands in 12/16; across 18 full-hierarchy trials, the selected specialist succeeded in 15/18 (83.3%). At runtime, evaluating the policy and both predictors averaged 0.29 ms on one CPU core, 1.45% of the 50 Hz control period.
The authors state that extending viability prediction to other tasks is future work, and thresholds are selected empirically and require tuning. The paper is a preprint and no code release is stated. Real-robot evaluation is limited to side flips on LimX Oli with small sample sizes (13 prediction trials, 18 full-controller trials, 19 abort and 16 protective-fall manual trials). Simulation uses two specific motions (front flip on G1, side flip on Oli) and randomized disturbances, but generalization to other maneuvers and platforms remains to be shown. Training relies on simulated switch-state distributions that may not cover all failure modes.
Humanoid robotics companies developing or deploying dynamic capabilities (e.g., Unitree, LimX Dynamics, Boston Dynamics, Figure) could integrate VAPS as a safety layer to reduce hardware damage during development, testing, and field operation of acrobatic or high-speed motions. It is also relevant to any legged-robot platform where a failed maneuver can destroy sensors or hands. The technique is demonstrated on real hardware now, but broad product adoption is likely 1–3 years away, requiring expansion to multiple maneuvers, validation across more platforms, integration with production control stacks, and safety certification. Before that, companies would need robust sim-to-real transfer for their specific robots and the ability to collect or generate failure data to train backup policies and predictors.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.