Auditing simulated robot failures traced 12 in 100 episodes to one physics-engine constant, and changing only that constant took the artifact rate to zero
AspireAI community · arXiv
Simulation benchmarks score robot policies by success rate, and every failed episode counts against the policy, even when the simulator caused the failure. Testing a vision-language-action model and a world action model, the authors found the policy often placed an object and the simulator then ejected it, and that the benchmark's own scripted demonstrations failed the same way. SimuGuard is an auditing plugin for LIBERO, RoboTwin, RoboCasa and RoboDojo that leaves the evaluation unchanged: at every physics step it records the state and flags an object that starts moving at a contact faster than robot contact or gravity can account for, then replays the episode bit for bit and reruns it with one simulation parameter changed. On RoboTwin the artifact episodes traced to the physics engine's collision detection: for geometry one to two centimetres apart the engine reports a penetration of about two centimetres and throws the object out at several metres per second, caused by one constant, a fixed-area threshold below which valid triangles are discarded. Changing that constant and nothing else dropped artifact episodes from 12 in 100 to none and raised success on five container tasks from 89% to 98%. The limitation is that this is one engine's artifact, and other simulators may have their own, so the audit method rather than the specific fix is the transferable result.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.