Scoring a robot policy on task success alone let a door-trained policy open drawers 59 times out of 110 while commanding 40 times the off-axis rotation a drawer-trained policy did
AspireAI community · arXiv
Manipulation policies are usually graded on binary task success, which cannot tell whether the commanded gesture is one the object's kinematics even admit. Constraint-Referenced Strategy projects the commanded end-effector twist onto the object's joint axis, separating real progress along the joint from rotation that goes nowhere, and it needs only a joint model and a scripted joint-following controller - no recorded demonstration and no learned judge.
Across 74 simulated articulated objects, success scores registered three transfer failures as lower rates but attributed none of them, while the two CRS channels gave each a distinct signature. A door-trained behaviour-cloning policy opened drawers in 59 of 110 episodes, and in those successful passes it commanded about 40 times the off-axis rotation of a drawer-trained policy's passes, 1.27 against 0.032 rad/s. The method is also comparatively robust to the joint model it depends on: verdicts are unchanged under joint-axis errors of up to 30 degrees, though absolute thresholds tolerate only about 8.
The limits are real. Everything is simulation, and a CRS verdict is only as good as the joint model, which the method assumes you have. The policy studied here is trained by behaviour cloning rather than reinforcement learning, and the three transfer failures the authors construct are their own. The authors present CRS as an attribution tool that says why a policy failed, not as a replacement for success scoring.
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.