Building task-specific simulations let a robot policy improve itself with 400 simulated trajectories and 10 real ones, beating 200 real demonstrations per task
AspireAI community · arXiv (cs.RO)
Adapting a robot manipulation policy to a new task is data-hungry, and the data that helps most depends on the policy's current weaknesses. EmbodiRSI runs recursive self-improvement in a real-to-sim-to-real loop, constructing task-specific simulations from the deployment scenario as a cheap workspace, then closing the loop with Collaborative Error Correction, which generates agent-assisted corrective trajectories from states the policy actually reached, and Adaptive Data Collection, which directs expert demonstration generation toward the current weaknesses. Across three tabletop environments and 14 subtasks it raised scene-balanced autonomous simulation success from 50.4 to 83.5 percent over two rounds, and with 400 adaptive simulated trajectories plus only ten real-world refinement trajectories per subtask it reached 83.1 percent scene-balanced autonomous real-world success, against 75.0 percent for adaptation using 200 real demonstrations per subtask. The limitation: the sim-to-real transfer is tied to simulations that can be built for the deployment scenario, which does not generalize to settings where you cannot model the target environment, and the real-world comparison uses a modest number of subtasks.
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.