How Two Trossen Arms Learned to Strike a Match
Two of our robot arms struck a match completely on their own. No one was teleoperating them. The arms ran a learned policy and carried out the task autonomously.
Striking a match looks simple, but it is one of the harder things you can ask a pair of robot arms to do. One arm has to hold the matchbox steady while the other brings the match across the striking surface at the right angle, with the right pressure, at the right moment. Push too softly and nothing happens. Push too hard or miss the timing and the match snaps or slips. Both arms have to work as one.
This demo comes from new research heading to the Conference on Robot Learning (CoRL) this November, and it was run on our hardware.
What NAC is, in plain terms
NAC treats robot movement the way music apps treat sound, and that makes it easier for robots to learn. It comes from "NAC: Neural Action Codec for Vision-Language-Action Models", first author work by Ahad Jawaid with Yu Xiang at The University of Texas at Dallas, accepted to CoRL 2026.
A robot moves by following a steady stream of tiny instructions, many per second, for every joint. That stream looks a lot like an audio signal, which is also a steady stream of tiny values over time. Music apps have spent years learning how to shrink audio into a small, compact form without losing what matters.
Ahad's idea was to borrow those same audio compression tools and point them at robot movement. The main change was dropping the parts built around how human ears hear sound, which don't apply to a robot's joints. The result packs each short chunk of movement into just 12 tokens, a compact code the robot's AI model can learn from and predict.
Result | NAC | Compared with |
Success rate, 8 real world tasks | 50% | 40% (FAST and OAT) |
Success rate, LIBERO-10 simulation | 49.7% | 44.2% (OAT) |
Tokens per action chunk | 12 | 12 (OAT and VQ-VLA), 36 (FAST), 224 (simple binning) |
The real world tasks included grasping grapes, weighing objects, stacking blocks and folding a towel, with 10 trials each (project page).
Why this matters for robot learning
The way a robot's movement gets encoded shapes how well it can learn. Many of today's robot AI models, known as vision language action models, think in tokens the same way chat models do. If movement takes too many tokens, the model is slower and has more room for error. If the encoding is too crude, fine details get lost and the robot fumbles delicate moves.
NAC's gains showed up most on tasks that need small, precise corrections, like grasping grapes and stacking blocks. Striking a match calls for that same kind of fine, coordinated control.
It is also a good reminder of how research moves forward. A proven idea from one field, audio, turned out to help with a hard problem in another. Research teams need hardware that lets them test ideas like this quickly on real robots, and that is exactly what we build for.
Built on our legacy arms, and there's more ahead
The match striking demo ran on our previous generation of arms. Our 2026 lineup is now live and built to take on even more dexterous work.
WidowX arms handle manipulation across research setups. Workbench and Rivet give teams bimanual platforms, stationary and mobile. Glide and Cockpit cover teleoperation and data collection, so teams can record demonstrations, train and deploy on one system.
We also offer academic discounts for universities and research labs. If you're planning your next setup, explore the 2026 lineup or reach out to our team.
Congratulations to Ahad and Yu on NAC. Read the paper, browse the code and see more results on the project page.
Comments