RAX learns robot tasks through online exploration

What happened

The researchers tested single-arm, two-arm, dexterous, mobile and humanoid systems across simulated tasks and physical robot experiments. They report a 37.4% average improvement over selected baselines in simulation; separately, they demonstrate real robots learning tasks such as articulated and long-horizon mobile manipulation through exploration without human demonstrations.

Why it matters

The study explores a route to adapting robots to new tasks without collecting a fresh teleoperation dataset for each one. The numerical benchmark gain is a simulation result, while the physical tests are controlled research demonstrations. The October 15 paper is a preprint, and its results have not been independently replicated or shown as a commercial deployment.

Source checked October 31, 2026. Status describes the evidence in the linked source at the time of review.