Problem
Research questions and friction points this paper is trying to address.
World Model
Representation Learning
Planning
Continuous Control
Embodied AI
Representation World Model: Learning States, Transition and Executable Plans in Representation
🤖 AI Summary
This work addresses the computational inefficiency of traditional world model planning, which relies on explicit dynamics modeling and search-based optimization. To overcome these limitations, we propose the "Representation World Model," which reformulates planning as geometric structure learning within a representation space. This approach jointly models states, transitions, and executable plans directly in the latent space, constructing latent trajectories through local inverse dynamics supervision. By eliminating recursive rollout and action-space search, it enables end-to-end direct planning. Experimental results demonstrate that the proposed framework achieves superior performance on continuous control benchmarks and exhibits strong potential for complex embodied tasks such as robotic manipulation.
📝 Abstract
We propose the Representation World Model (RWM), which learns states, transitions, and executable plans directly in representation space. Unlike existing world models that typically learn latent representations together with explicit dynamics models and perform planning through search, optimization, or policy-based prediction, RWM directly incorporates planning into the learned representation geometry. RWM learns the representation geometry by applying inverse-dynamics supervision locally along latent paths constructed from endpoint representations, requiring these paths to preserve task-relevant state and transition information. At inference, planning is performed by directly constructing a latent path between the current and goal representations, with inverse dynamics used to recover the corresponding actions, without recursive rollouts or action-space search. Experiments on continuous-control benchmarks demonstrate the effectiveness of RWM for direct planning, while results on robotic manipulation further show its potential to extend to more complex embodied control tasks. These results suggest that planning directly in representation space provides a promising alternative to conventional world-model planning.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Representation World Model
Latent Planning
Inverse Dynamics
Representation Geometry
Continuous Control
🔎 Similar Papers
No similar papers found.
