Search papers, labs, and topics across Lattice.
The paper introduces RISE, a framework for robotic reinforcement learning that uses a Compositional World Model to improve policy robustness in contact-rich manipulation tasks. RISE predicts multi-view futures using a controllable dynamics model and evaluates outcomes with a progress value model to generate advantages for policy improvement. The framework operates in a closed-loop, self-improving pipeline, updating the policy in imaginary space based on imaginary rollouts, achieving significant performance gains in real-world tasks compared to existing methods.
Robots can now learn complex manipulation skills entirely in simulation, thanks to a compositional world model that accurately predicts future states and evaluates progress, leading to a 35-45% performance boost in real-world tasks.
Despite the sustained scaling on model capacity and data acquisition, Vision-Language-Action (VLA) models remain brittle in contact-rich and dynamic manipulation tasks, where minor execution deviations can compound into failures. While reinforcement learning (RL) offers a principled path to robustness, on-policy RL in the physical world is constrained by safety risk, hardware cost, and environment reset. To bridge this gap, we present RISE, a scalable framework of robotic reinforcement learning via imagination. At its core is a Compositional World Model that (i) predicts multi-view future via a controllable dynamics model, and (ii) evaluates imagined outcomes with a progress value model, producing informative advantages for the policy improvement. Such compositional design allows state and value to be tailored by best-suited yet distinct architectures and objectives. These components are integrated into a closed-loop self-improving pipeline that continuously generates imaginary rollouts, estimates advantages, and updates the policy in imaginary space without costly physical interaction. Across three challenging real-world tasks, RISE yields significant improvement over prior art, with more than +35% absolute performance increase in dynamic brick sorting, +45% for backpack packing, and +35% for box closing, respectively.
