ForceHarness lifts frozen-model robot manipulation success from 3.3% to 90%
showlab · hf · 2026-11-18
Show Lab (NUS) presents ForceHarness, letting API-only frozen multimodal models use contact forces for real-robot manipulation without demos or finetuning.
- Challenges: Contact evolves faster than API decisions, requiring force closed-loop below the model; frozen models can't read raw force streams, so forces are translated to physical quantities with uncertainty.
- Approach: Hand-force calibration from wrist sensors or joint torques; self-terminating force primitives; a physics translation layer returning quantities as tool results.
- Results: On a real arm without a wrist sensor, GPT-6 Astra's zero-shot success jumps from 3.3% to 90.0%; in simulation, +17.5pp on ForceBench and +34.3pp on HiddenPhys (a new 7-task MuJoCo benchmark), beating demo-trained force-aware policies and matching expert hand-written controllers.
- Ablation: Smaller models decide much better given physical quantities over raw force values.
More from Embodied
- HubLabs' HubEgo-100K robotics dataset trends on Hugging Face — HubLabs · 2026-11-18
- ResWAM: residual learning on frozen DINOv3 boosts VLA success with less compute — OPTML-Group · 2026-11-18
- ImpedanceMimic learns per-joint impedance as action space, beating PHC and SimPoE on force-accurate motion imitation — tweetsatpreet · 2026-11-18
- Google building new robotics model Astra with 1M-token context, config leaks — lyraxana · 2026-11-18
- GPT-6.2 Astra tops robot-code test: 2.5B net solving 79% of LIBERO, Python code hits 96.8% — burny_tech · 2026-11-18
- Bittensor subnet's zels decision model 'Jev' integrated into Even Realities glasses — markjeffrey · 2026-11-18