Simulation-trained 27B model transfers robot skills to real dual arms
First merged summary.
A 27B reasoning model trained only in simulation succeeded in 60% of real dual-arm trials; the untuned base model succeeded in none.
Researchers trained an open-weight 27B reasoning model with robot-use reinforcement learning entirely in simulation, using a Post-training Engine for Agentic Robotics and no demonstrations. The model predicts the next end-effector pose from observations and chain-of-thought, without a harness. Skills transferred to an unseen robot, the real world, and novel instructions. After LIBERO and RoboDojo training, real dual-arm trials succeeded 60% of the time. On LIBERO, success rose from 1.6% to 74.4%, near GPT-6 Astra and GPT-6.1 Sol at maximum reasoning effort, with no measured drop on nine general benchmarks.
- Real dual-arm success was 60% after simulation-only training.
- LIBERO success rose from 1.6% to 74.4%.
- LIBERO-Pro success was 57.4% on a new task in a familiar scene.
- Demonstration-tuned π0.5 succeeded in 0.75% of those LIBERO-Pro episodes.
- Authors report no measurable drop on nine reasoning, coding, and vision benchmarks.
Timeline
dated events in the reporting · oldest first- Nov 13, 2026Paper reports simulation-only training and real-robot transfer. · arXiv
What each outlet adds
beyond the shared facts- arXivReal dual-arm rate, LIBERO gains, and no benchmark drop.
- Vendors
- DeepSeek
Coverage
oldest first · one row per article · UTC27B reasoning model learns real-robot control from simulation alone
It succeeded in 60% of real dual-arm trials after simulation-only training; the untuned base model succeeded in none.
ScaleVLA ties a 552B VLM backbone to a robot action model
With 10 demonstrations it beat a 2B backbone trained on 50, and LIBERO-PRO success was 45.1% versus at most 31.3%.