PhysClaw-0: Robot Data Collection With Reusable Language Corrections

PhysClaw-0 stores human language corrections as reusable robot adjustments, reducing repeated oversight during real-world manipulation data collection.

TL;DR — PhysClaw-0 is a real-robot data-collection system that stores human language corrections and reuses them across rounds. On a desktop-clearing testbed, it matches teleoperation episode success while reducing human working time to 16%.

The bottleneck: robots need data, and data needs people

Training robots to manipulate real objects requires real-world trajectories: records of what the robot sees, decides, and does. For tasks such as clearing a desktop, those trajectories must include attempts, failures, recoveries, and resets.

Collecting that data is costly because robots still need supervision. A person may have to reset the scene, judge whether an attempt succeeded, or intervene after a mistake. Even semi-autonomous pipelines can make people repeat the same fix whenever the same failure returns.

The paper “PhysClaw-0: A Symbiotic Agentic System for Robot Autonomy via Language Corrections” targets that repeated-oversight problem. Its core idea: when a person corrects the robot in natural language, the system should retain that correction and reuse it in later rounds.

Full analysis is available via x402 micropayment.