{"name":"RoboHarness","repo":"BinceQu/RoboHarness","url":"https://midorreal.com/project/bincequ-roboharness","repository":"https://github.com/BinceQu/RoboHarness","description":"A robotic harness that gives LLM agents a visual-geometric control panel so that they can directly understand and invoke embodied tasks.","facts":{"license":{"value":"MIT","source":"https://github.com/BinceQu/RoboHarness","checked":"22 Nov 2026"},"language":{"value":"Python","source":"https://github.com/BinceQu/RoboHarness","checked":"22 Nov 2026"},"homepage":{"value":"https://roboharness-lab.github.io/","source":"https://roboharness-lab.github.io/","checked":"22 Nov 2026"},"releases":{"value":"0","source":"https://github.com/BinceQu/RoboHarness/releases","checked":"19 Nov 2026"},"contributors":{"value":"2","source":"https://github.com/BinceQu/RoboHarness/graphs/contributors","checked":"19 Nov 2026"},"archived":{"value":"no","source":"https://github.com/BinceQu/RoboHarness","checked":"22 Nov 2026"}},"stars":539.0,"writeup":{"what_it_is":"RoboHarness is a harness that gives LLM agents a visual-geometric control panel for robots. The LLM marks points on a 2D image, and the system tracks them with optical flow and depth back-projection to compose primitive actions for manipulation tasks. It includes a V1 Claude Code/Codex plugin setup and a V2 workspace harness, and is evaluated on the BEHAVIOR simulator.","audience":"Researchers and developers evaluating LLM agents for robotic manipulation in simulation","claims":[{"kind":"broad","claim":"Outperforms VLA and world action models","excerpt":"A Simple Harness Outperforms VLA and World Action Models","status":"too_broad"},{"kind":"specific","claim":"Uses five evaluation instances per task with fixed instance IDs","excerpt":"Every task uses evaluation slots **0, 3, 5, 7, 9**","status":"not_checked"},{"kind":"specific","claim":"Requires Linux x86-64, an NVIDIA RTX GPU, Python 3.11 and CUDA 12.4","excerpt":"Use Linux x86-64, an NVIDIA RTX GPU, Python 3.11 and the CUDA 12.4 toolkit.","status":"not_checked"},{"kind":"specific","claim":"Provides a V2 workspace harness that runs with --harness v2","excerpt":"It runs on the same tasks and evaluator with `--harness v2`","status":"not_checked"},{"kind":"specific","claim":"Tracks visual keypoints persistently and verifies actions in a closed loop","excerpt":"visual keypoints are tracked persistently, converted into geometric constraints, executed, and verified in a closed loop","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-12"},"why_now":null,"owner_supplied":null,"score":{"verdict":"mid","hype":53,"reality":47,"gap":7,"momentum":0.3,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-10-07T04:39:45+00:00","text":"Repository created","source":"https://github.com/BinceQu/RoboHarness"}],"cite":"RoboHarness is a GitHub project at https://github.com/BinceQu/RoboHarness written mainly in Python, described there as \"A robotic harness that gives LLM agents a visual-geometric control panel so that they can directly understand and invoke embodied tasks\".","attribution":"Data from Mid or Real, https://midorreal.com/project/bincequ-roboharness, checked 22 November 2026. Attribution with a link is required.","still_real":null}