huggingface.co web signal

NUS ForceHarness Lifts GPT-6 Astra Real-Robot Success From 3.3% to 90% on Force-Rich Manipulation, No Demos

Robotics OpenAI Multimodal ai-business

Summary

Singapore's Show Lab released ForceHarness, a middleware layer letting frozen API-only models like GPT-6 Astra run force-sensitive manipulation with no fine-tuning or demonstrations. On a Franka FR3 with no wrist sensor, real-robot success on six contact-rich tasks jumped from 3.3% to 90.0%. The system layers 12 self-terminating force primitives and a physics-translation stage that converts raw force into calibrated mass, friction and contact-height estimates.