Panasonic T-Rex robot AI hits 65% success on contact-rich bimanual tasks with fingertip touch
VLA action generation at ~5 Hz plus tactile correction at ~20 Hz beat prior methods by 30 points across 12 real tasks; CoRL 2026 presentation with Berkeley, NVIDIA and Stanford.
Why it matters
Most vision-language-action robots still struggle when force, slip or soft-object deformation matter. T-Rex keeps a slow vision-language planner and overlays a faster fingertip loop so the hands can correct without re-running expensive image inference. On tasks such as transferring eggs, unlocking keys and screwing bulbs it averaged 65% success versus 35% for the prior best baseline, and 42% without tactile input. The training mix includes over 7,700 bimanual episodes plus large egocentric video pretraining. Watch whether the CoRL paper releases code or hardware details that others can reproduce outside Panasonic’s lab setup.
Key facts
- 65% average success across 12 contact-rich real tasks
- Prior best baseline 35%; without tactile input 42%
- Vision-language actions ~5 Hz; tactile correction ~20 Hz
- 7,700+ bimanual episodes (~100 hours) with 10 fingertip sensors
- Co-developed with UC Berkeley, NVIDIA, Stanford and others
- To be presented at CoRL 2026 in Austin, 9–12 Nov