Reproduce “Steering Robots with Inference-Time Interactions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yanwei Wang

Yanwei Wang's doctoral thesis from MIT introduces methods for steering robot behavior at inference time using real-time user interactions, moving beyond costly retraining. The research proposes two frameworks, ITPS for flexible probabilistic guidance and TAMI for robust, constraint-satisfying multi-step tasks, demonstrating how to achieve reliable robotic performance in dynamic, user-involved environments.

Key insights

  • A critical trade-off exists between aligning robot behavior with user intent and maintaining the policy's distribution to avoid failures; stochastic sampling effectively balances this.
  • Formal bisimulation properties, specifically reachability and invariance within task modes, are essential to guarantee task success and prevent discrete-level looping in multi-step robotic operations under perturbation.
  • Semantic mode boundaries and task constraints can be robustly learned from limited demonstrations and synthetically generated counterfactual data, guided by large language models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.