Robots Rehearse Delicate Work Before the Factory Floor

The story of robot learning used to be locomotion — walking, hopping, recovering from a shove. The work presented at the Conference on Robot Learning (CoRL) in Austin this week points somewhere harder: manipulation that needs a finer touch, like aligning gears, routing cables, and assembling millimeter-scale parts.NVIDIA used the conference to introduce GR00T 2, an open world action model; Isaac Lab 3.0; and open industrial benchmarks built with partners S1. Days earlier, the Allen Institute for AI published ToyBlocks, a deliberately small simulated world built for the kind of controlled experiments the industrial push needs S2. Read together, they describe a maturing workflow: pretrain, rehearse in physics, evaluate across variations, then ship to hardware.

From walking to cable swaps

GR00T 2, trained on diverse robot and human data including 100,000 hours of open human egocentric data, completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations, reaching 71% average success with one demonstration per task — 95% with five S1. Foxconn is applying it to server manufacturing and Siemens is evaluating it across industrial use cases. Underneath, Isaac Lab 3.0's tighter integration with the Newton physics engine lets simulators model deformables such as cables and cloth, which is precisely what assembly work demands. On the NIST Assembly Task Board, a single policy handled all 20 tasks with more than 98% training success versus 81% for separate per-task policies, using roughly 39 times less GPU time; behind that result is Success-Guided Sampling, which directs practice toward skills a policy only solves about half the time S1.

Sim-to-real numbers, reported on both sides

The most encouraging detail in AI2's ToyBlocks release is the accounting. Policies trained purely in simulation ran zero-shot on a real Franka FR3 arm across thirty scenes: the ManiFlow low-level model scored 80.0% in simulation and 81.1% in reality on individual skills, while the hierarchical Molmo 2 plus ManiFlow system managed 55.5% simulated versus 50.0% real on multi-step block tasks — with the authors noting that further trials are needed for statistical confidence S2. Their generalization findings are a caution: despite large-scale pretraining, popular action models still generalized poorly in the initial results, particularly beyond the spatial boundaries of their training data. The same testbed showed that adding 100 model-generated rollouts from the best video world model lifted a policy's success in an unseen region from 19% to 78%. Industrial evaluation is formalizing in parallel, with DexBench's 15 tasks spanning retail, logistics, manufacturing and hospitality, UC Berkeley's six-task VABAR, and the Isaac Lab-Arena framework S1. LG Electronics demonstrated GR00T-powered manipulation on its CLOiD robot in an actual manufacturing environment, wiring real-world data collection, training and simulation-based evaluation into one repeatable loop S3.

Agents are becoming robot programmers

The third shift is who writes the robot program. UC Berkeley's Graph-as-Policy framework uses agents to chain individual skills into robot programs that stay interpretable, so engineers can review and verify them before deployment S1. AI2 exposed ToyBlocks skills to a Claude Opus 5.5 agent over MCP: it solved or correctly declared infeasible 63% of search-and-spell tasks, and skill descriptions the agent wrote itself lifted success from 24% to 63%, beating the 36% achieved with descriptions written by a human expert S2. The failures are just as instructive — mostly look-alike letters and rotated orientations in block stacks. The practical takeaway for deployment teams: let agents compose and even improve robot skills, but keep a human review gate, because a planner that misreads the scene will confidently build the wrong tower.