Rollback World Model Beats RISE 24%, Tops Humans in 8 Hours

A policy improvement method called Rollback reports a 24.0% relative gain over the RISE with GE-Sim 2.0 baseline across three real-world robot tasks, and surpasses human average success on two of them within 8 hours of training, according to a research team's paper posted on Hugging Face.

The core idea reframes how learned simulators get used. "The world model becomes not merely a surrogate simulator, but a rewindable training environment," the authors write. Rather than roll forward and discard trajectories, Rollback branches at critical decision points, contrasts counterfactual outcomes, and distills them into action-level supervision without a separately learned value model.

On Block Extraction (pulling Jenga blocks without collapsing the tower), the paper reports 69.33% success, a 10.67 point absolute gain over RISE with GE-Sim 2.0. Ball Tossing hits 40.00% against a 31.48% human average; Tabletop Curling reaches 64.44% against 58.39% for humans. The run uses 40 NVIDIA H200 GPUs for tree search and 8 more for policy training, on top of a π₀.₅ vision-language-action policy.

The paper also pairs the search loop with test-time grounding: "real transitions from the policy it trains correct its dynamics through a lightweight LoRA pipeline." The study benchmarks three tabletop tasks; transfer beyond them is a question for future work. The result lands amid steady movement on the robot policy learning beat, with our research tracker logging 43 papers in the last 90 days.