Robot arm teaches itself to rack vials: one GPU, 27 min, 1010/1024 success
MaziyarPanahi · x · 2026-11-21
A hobbyist got a robot arm to teach itself to rack specimen vials using RL. Training ran across 4,096 parallel simulated benches and took just 27 minutes on a single GPU.
The learning curve is striking:
- 0 successful racks at iteration 200
- 24 at iteration 250
- 843 at iteration 300
- 1,010 out of 1,024 tries at the end (98.6% success)
As the author put it: 200 rounds of nothing, then it clicked — a textbook example of RL's sudden breakthrough-style convergence.
More from Embodied
- Wayve self-driving demo handles complex Tokyo streets with zero interventions — alexgkendall · 2026-11-21
- Robot can restock the grocery store? Employees post chaotic photo in reply — MatthewChang · 2026-11-21
- FOSDEM 2027 Robotics & Simulation devroom opens call for talks, deadline Nov 23 — facontidavide · 2026-11-21
- Google's VERID grounds AI agents in wet labs, finds nanomolar cancer protein binders — CurieuxExplorer · 2026-11-21
- SuperMimic took 11 months of code and experiments; sim-to-real may move past PPO — ChongZzZhang · 2026-11-21
- Autonomous cars found draining batteries overnight by constantly reparking themselves — mtizard · 2026-11-21