Continue, Abort, or Fall: Viability-Aware Policy Selection (VAPS) for Safe Humanoid Acrobatics

Siwei Ju; Lu Liu; Jan Peters; Oleg Arenz

cs.ROOct 1, 2026

Continue, Abort, or Fall: Viability-Aware Policy Selection (VAPS) for Safe Humanoid Acrobatics

Authors: Siwei Ju, Lu Liu, Jan Peters, Oleg Arenz

Organizations: Department of Computer Science, Technical University of Darmstadt, Germany. · Robotics Institute Germany (RIG). · LimX Dynamics. · German Research Center for AI (DFKI), Research Department: Systems AI for Robot Learning.

Abstract

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.

Figures & tables

Fig. 2 : Overview of VAPS . (a) Specialist policy training. (b) Receding-horizon predictor training. (c) VAPS : The least-sacrificial rule of policy selection.

Policy	DR	Specialist Success	Head contact	Hand contact
Nominal	1	$100.0\pm 0.0\%$	$0.0\pm 0.0\%$	$0.0\pm 0.0\%$
	2	$86.2\pm 1.6\%$	$8.1\pm 1.1\%$	$13.6\pm 1.6\%$
	4	$15.0\pm 2.2\%$	$62.0\pm 2.9\%$	$82.3\pm 2.0\%$
Abort	1	$100.0\pm 0.0\%$	$0.0\pm 0.0\%$	$5.0\pm 0.0\%$
	2	$89.0\pm 8.2\%$	$9.0\pm 6.5\%$	$11.0\pm 7.4\%$
	4	$62.0\pm 13.0\%$	$17.0\pm 8.4\%$	$26.0\pm 6.5\%$

TABLE I : Specialist performance across domain randomization (DR) scales. Specialist success: Nominal – motion success, Abort – upright stance, ProtFall – head clearance.

Method	Episode Acc. $\uparrow$	Detection $\uparrow$	FAR $\downarrow$	LT mean $\uparrow$
SafeFall ( $t_{2}=$ 100\text{,}\mathrm{m}\mathrm{s}$$ )	$95.5\pm 2.1\%$	$85.8\pm 7.0\%$	$0.4\pm 0.8\%$	$0.205\pm 0.012$ s
SafeFall ( $t_{2}=$ 200\text{,}\mathrm{m}\mathrm{s}$$ )	$95.0\pm 1.5\%$	$83.3\pm 5.1\%$	$\mathbf{0.0\pm 0.0}\%$	$0.259\pm 0.014$ s
SafeFall ( $t_{2}=$ 500\text{,}\mathrm{m}\mathrm{s}$$ )	$95.0\pm 0.0\%$	$84.2\pm 1.9\%$	$0.4\pm 0.8\%$	$0.316\pm 0.063$ s
SafeFall ( $t_{2}=$ 1000\text{,}\mathrm{m}\mathrm{s}$$ )	$93.5\pm 1.6\%$	$79.2\pm 5.1\%$	$0.4\pm 0.8\%$	$0.281\pm 0.046$ s
VAPS ( $\gamma=0.9,H=50$ )	$\mathbf{97.5\pm 1.5}\%$	$\mathbf{91.7\pm 5.1}\%$	$\mathbf{0.0\pm 0.0}\%$	$\mathbf{0.416\pm 0.042}$ s

TABLE II : Predictor comparison against SafeFall-style final-outcome baselines. FAR: false alarm rate. LT: lead time.

Fig. 3 : Predictor output for two representative failed real-robot side flips. The dashed line marks take-off.

	All episodes (%)					Fallen episodes only, median [mean]
Method	Success $\uparrow$	Head $\downarrow$	Hand $\downarrow$	Other $\downarrow$	Fall $\downarrow$	Pk. head (N)	Pk. non-foot (N)
G1, Nominal only	62.1 $\pm$ 0.2	27.7 $\pm$ 0.2	36.6 $\pm$ 0.2	36.3 $\pm$ 0.1	37.9 $\pm$ 0.2	540.2 $\pm$ 13.8 [954.8 $\pm$ 122.0]	2124.6 $\pm$ 33.9 [2608.6 $\pm$ 352.2]
G1, VAPS — shared alarm rule, $\tau_{\mathrm{nom}}=0.20$
VAPS (Nominal–Abort)	59.6 $\pm$ 0.4	5.0 $\pm$ 0.4	7.4 $\pm$ 0.5	5.5 $\pm$ 0.3	7.4 $\pm$ 0.5	833.5 $\pm$ 139.1 [1076.0 $\pm$ 72.1]	2219.5 $\pm$ 87.2 [2110.4 $\pm$ 189.7]
VAPS (Nominal–ProtFall)	59.6 $\pm$ 0.4	2.0 $\pm$ 0.1	11.8 $\pm$ 0.6	39.9 $\pm$ 0.5	40.4 $\pm$ 0.4	0.0 $\pm$ 0.0 [21.8 $\pm$ 2.6]	1871.8 $\pm$ 14.2 [2058.6 $\pm$ 17.4]
VAPS (one-shot)	59.6 $\pm$ 0.4	2.2 $\pm$ 0.2	9.3 $\pm$ 0.4	25.6 $\pm$ 1.7	26.2 $\pm$ 1.7	0.0 $\pm$ 0.0 [62.2 $\pm$ 12.9]	1855.3 $\pm$ 38.1 [2073.4 $\pm$ 77.2]

TABLE III : Closed-loop performance under randomized disturbances.

\pm

is over five seeds. Results are on Unitree G1 except for the last two rows.

Fig. 4 : Head-contact rate against motion success on the Unitree G1 results of Table III ; down and right is better.

Fig. 5 : VAPS applied to

16

checkpoints of an independently trained nominal policy.

Explore similar work

May 18, 2026cs.RO

Unified Walking, Running, and Recovery for Humanoids via State-Dependent Adversarial Motion Priors

We propose a unified reinforcement learning framework that enables a single policy to perform walking, running, and fall recovery on the Unitree G1 humanoid robot, validated on physical hardware without any explicit mode-switching command at deployment. The framework extends Adversarial Motion Priors (AMP) by replacing the conventional global reference distribution with a state-dependent gate that routes each training transition to one of two discriminators: a dedicated recovery discriminator and a velocity-conditioned locomotion discriminator that jointly covers walking and running. The gate is defined by a single fixed threshold on projected gravity: the recovery discriminator is activated when body tilt exceeds approximately

37^\circ

from vertical (

|g_z+1|>0.6

); otherwise the locomotion discriminator is used, with the normalized commanded velocity serving as a condition that selects the appropriate reference trajectory between walk and run clips. Only three LAFAN1 reference clips are required to regularize the complete behavior set. At deployment, a single frozen ONNX policy executes at 50,Hz with no runtime mode logic; hardware experiments demonstrate successful recovery from both prone and supine falls and smooth walk-to-run transitions under the same controller.

Yidan Lu, Yichao Zhong, Liu Zhao +2

The University of Hong Kong, Hong Kong SAR, China.

Full-Body Humanoid Control Offline Reinforcement Learning Humanoid

Ask PDF

Sep 14, 2026cs.RO

ResSafe: Learning Safety Filtering with Residual Reinforcement Learning for Humanoids

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

Safety Filters Full-Body Humanoid Control Residual Policy Humanoid Nonlinear Dynamics Pareto Frontier

Ask PDF

Sep 22, 2026cs.RO

SafeLoop: Risk-Aware Rollback for Vision-Language-Action Manipulation

Recent vision-language-action (VLA) models are promising for general-purpose manipulation, but long-horizon execution remains fragile. Small state-estimation or control errors can lead to irreversible failures (e.g., collisions and object drops). Avoiding these risks requires a proactive safety mechanism capable of anticipating hazards. In this paper, we introduce SafeLoop, a non-invasive external wrapper that adds hazard prediction and rollback-based recovery to a VLA model without changing its parameters. SafeLoop trains a risk predictor from vision and proprioception to output four values: the probability and time-to-hazard for body collisions and for object failures. A lightweight controller then chooses one of three actions based on the predicted risk: continue execution (noop), save a safety checkpoint (record), or retreat in joint space (rollback). Rollback moves the robot back to a recent safe waypoint and queries the base policy again, which may yield an alternative continuation. Across 24 LIBERO tasks (16 random seeds each) and three real-robot tasks (25 rollouts each), SafeLoop achieves a stronger overall safety-success trade-off than alternative methods, reducing hazard cases by roughly 70% while preserving task success and the base-policy control rate. Project code is available at https://github.com/Loule0-0/SafeLoop/tree/release/safeloop.

Zeyu Lou, Tianran Zhang, Xinquan Yue +2

Nanjing University, Nanjing, China · The Hong Kong University of Science and Technology (Guangzhou), Guangdong, China · Beijing University of Technology, Beijing, China

Vision-Language-Action Framework Robotic Manipulation Safer Hazard Rollback

Ask PDF