Global Tech News For people figuring out what to build with AI.
Robotics

Guided training cut human corrections by 82% in a bread-insertion task

By ·

Diagram of RoboPrompt turning a human visual or directional prompt into an action draft, then refining it with a robot policy.

A prompt-steering system let people guide a robot policy with a drawn path, target point or broad direction. After two or three rounds of guided training, the researchers report 81.9% fewer human steering interventions on a physical bread-insertion task, averaged across three robot policies, from 2.60 to 0.47.

The system, RoboPrompt, addresses a practical gap: learned robot policies can complete tasks, but users have few simple ways to correct them when they begin to go wrong. A separate 0.77-billion-parameter vision-language-action model translates a sketch, point or short direction into an action draft. The robot’s existing diffusion or flow-matching policy then refines that draft, balancing the person’s instruction with what the policy has learned. The interface does not require changing the base policy architecture or tuning it specifically for steerability.

The team tested RoboPrompt with Diffusion Policy, π0.5 and FastWAM on a Flexiv Rizon4 robot arm. The prompts aligned with the requested movement in more than 95% of evaluated interventions for each policy, under the paper’s measure of prompt alignment. The researchers then used guided runs to train the policies over two or three Dataset Aggregation (DAgger) rounds, an iterative method that adds corrected examples to training. For π0.5, the paper reports a 15.5% average success-rate increase across bread insertion, hanging a cup and pushing a ball. The 81.9% reduction was measured on bread insertion across all three policies.

That result comes from a small laboratory evaluation, not a broad deployment test. Each task used 100 teleoperated demonstrations, and the experiments used one robot-arm platform. Six first-time volunteers received a five-minute introduction and each completed five bread-insertion trials, a limited usability check. The authors say RoboPrompt currently steers an end-effector and does not yet control dexterous multi-finger hands. The work is a new arXiv preprint.

Sources