General purpose multimodal models now directly control real robotic arms. Performance on a six-task benchmark doubled every 53.9 days from July to October 2026. The best execution rate rose 4.4 times while the lowest cost fell 6.2 times. Models rarely get stuck but often misjudge grasps or distances. Stronger models diagnose and compensate for physical imperfections better. Efficiency, not reasoning, is the main current limiter. The work uses a bimanual YAM robot with wrist and overhead cameras. Nine models were tested across fifteen configurations. A paper with full methodology was submitted to TAG@CoRL.
What changed
General purpose multimodal models were poor at controlling robotic arms in 2025.
What it unlocks
Directly controlling real robotic arms to manipulate physical environments with steady progress.
- 53.9-day performance doubling time
- 4.4x best execution rate increase
- 6.2x lowest execution cost reduction
- 45-48% mean task scores for strongest models
What you need to act on it
- frontier multimodal language models
- real robotic arms with proprioceptive feedback and camera footage
Sources