CoRL results show sim-only training now delivers top real-world robot performance
carlosdponx · x · 2026-11-12
Recent CoRL papers suggest simulation is the key to scaling robotics, with several sim-trained systems posting standout real-world results:
- MolmoBot: trained only in MuJoCo simulation, 79.2% zero-shot on real tasks
- GRAIL: humanoid data generated from 3D assets + a video model (no teleoperation); 84% pick-up on a real G1, 90% on stairs
- ZAP: open 27B model trained purely in simulation, 96.7% on a real bimanual arm at 1/17 the cost of a frontier model
- RPG: improved from 28.6% to 95% within 15 practice rounds
- DreamZero: 12 minutes of human video lifted unseen-task performance from 38% to 54%
- GR00T 2: 95% on new tasks from just 5 demos, with training mostly from human video
The takeaway: the best real-world robot results are now coming from robots that never touched the real world.
More from Embodied
- Baidu World 2026 to showcase Apollo Go robotaxi rides and Apollo World physical AI model — Baidu_Inc · 2026-11-12
- NVIDIA's open-source Alpamayo 2 Edge VLA driving model lands in Autoware — drmapavone · 2026-11-12
- A closer look at Astra's robot gripper design (and a joke at its expense) — _Stocko_ · 2026-11-12
- Robot startup passes first presale milestone with 10+ units claimed — R4nsack · 2026-11-12
- Runway launches Praxis-1, a 27B MoE world action model, with a 73ms Fast variant — ChongZzZhang · 2026-11-12
- Dream-1 on bimanual YAM robot folds a towel on voice command at CoRL — DJiafei · 2026-11-12