bincequ/roboharness is a robotics harness that gives LLM agents a visual-geometric control panel. The agent marks points of interest in a 2D image. It follows those points via optical flow tracking, receives positions through depth back-projection, and then composes primitive actions to perform manipulation. Rather than training a new model, it combines visual clicking, keypoint tracking, and geometric constraints. It is written in Python and released under the MIT license.
Instead of relying on a trained vision-language-action model, it takes the approach of having the agent receive coordinates directly and assemble primitive actions. The repository contains the BEHAVIOR simulator and evaluator, an RGB-D interface, Claude Code and Codex harnesses, task definitions, and an evaluation pipeline, so the required actions and pass criteria can be reviewed together. A reproduction contract is also published: it specifies fixed seeds, five instances per task with fixed instance numbers, the official evaluator, and tolerances, and it includes verification artifacts and baseline results in the repository. The repository was created on October 7, 2026, and has 250 stars.
However, there is no automated check workflow. The full run has to be executed by a person, and the actual comparison figures depend on the local model server. There are no registered releases either. In the collected evaluation record, the scores are novelty 7.0, tech 7.0, quality 7.0, and momentum 1.2, so only momentum is low. The remaining limitation is that comparison figures vary with the local model server and environment configuration.