Tsinghua, NYU and Northwestern unveil RoboDoctor, a training-free framework for fixing robot actions

jiqizhixin · x · 2026-11-05

Researchers from Tsinghua University, NYU and Northwestern University introduce RoboDoctor, a training-free execution framework that keeps both policy and agent weights fixed.

The core problem: when a robot action looks nearly right, should the system preserve it, correct a small part, or take over? Repeated intervention can disrupt useful motion, and an action proposal alone doesn't confirm the intended physical effect actually occurred.

How it works (the figure shows a three-stage RSI pipeline):

A warmup phase (3 hybrid episodes) chooses between policy-led hybrid control and agent-led direct control based on mean native score;
A Doctor module inspects observations and 32-action chunk previews, retaining useful prefixes (≤16 unchanged steps) or issuing 1–5 grounded corrective steps;
The operation stage follows ground → specify method → execute → verify effects, then revises from feedback;
An external experience store logs diagnoses, operations and evidence (e.g. "check alignment before tilt") for reuse across scenes.

Evaluated on a curated 54-task simulation archive across five dimensions, the framework improves execution reliability without touching model weights.

Original post →

Tsinghua, NYU and Northwestern unveil RoboDoctor, a training-free framework for fixing robot actions

More from Embodied