{"paper":{"title":"Prismatic World Model: Learning Compositional Dynamics for Planning in Hybrid Systems","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A context-aware mixture of experts decomposes hybrid robot dynamics into distinct modes to reduce rollout drift.","cross_cats":["cs.RO"],"primary_cat":"cs.AI","authors_text":"Chengwei Yang, Mingwei Li, Xiaoyuan Zhang, Yaodong Yang, Zilong Zheng","submitted_at":"2025-12-09T09:40:34Z","abstract_excerpt":"Model-based planning in robotic domains is challenged by the hybrid nature of physical dynamics, where continuous motion is punctuated by discrete events such as contacts and impacts. Conventional latent world models typically employ monolithic neural networks that enforce global continuity, which over-smooths distinct dynamic modes (e.g., sticking vs. sliding, flight vs. stance). For a planner, this smoothing results in compounding errors during long-horizon lookaheads, rendering the search process unreliable at physical boundaries. To address this, we introduce the Prismatic World Model (PRI"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"By modeling the mode transitions in system dynamics, PRISM-WM reduces rollout drift. Experiments on continuous control benchmarks, including high-dimensional humanoids and multi-task settings, demonstrate that PRISM-WM provides a high-fidelity substrate for trajectory optimization algorithms (e.g., TD-MPC).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That an implicit gating mechanism can reliably identify distinct physical modes from context alone and that the latent orthogonalization objective will prevent mode collapse without explicit mode labels or additional regularization.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PRISM-WM uses a context-aware MoE with latent orthogonalization to model hybrid dynamics and reduce rollout drift for model-based planning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A context-aware mixture of experts decomposes hybrid robot dynamics into distinct modes to reduce rollout drift.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"24e744e32e21649e5892bc34515b3470b3b29d07d429d3dfa4e8e67df690df19"},"source":{"id":"2512.08411","kind":"arxiv","version":2},"verdict":{"id":"5e1f8b65-7fd4-4484-b797-de9b918d6c7d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T00:32:12.664093Z","strongest_claim":"By modeling the mode transitions in system dynamics, PRISM-WM reduces rollout drift. Experiments on continuous control benchmarks, including high-dimensional humanoids and multi-task settings, demonstrate that PRISM-WM provides a high-fidelity substrate for trajectory optimization algorithms (e.g., TD-MPC).","one_line_summary":"PRISM-WM uses a context-aware MoE with latent orthogonalization to model hybrid dynamics and reduce rollout drift for model-based planning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That an implicit gating mechanism can reliably identify distinct physical modes from context alone and that the latent orthogonalization objective will prevent mode collapse without explicit mode labels or additional regularization.","pith_extraction_headline":"A context-aware mixture of experts decomposes hybrid robot dynamics into distinct modes to reduce rollout drift."},"references":{"count":5,"sample":[{"doi":"","year":1912,"title":"Soft Actor-Critic Algorithms and Applications","work_id":"bb49c9fb-03b2-4226-9edb-50186b8193e4","ref_index":1,"cited_arxiv_id":"1812.05905","is_internal_anchor":true},{"doi":"","year":2023,"title":"InAdvances in Neural Information Pro- cessing Systems, volume 28, 2944–2952","work_id":"efa01e2d-1160-4df3-a516-3972d2f961e2","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Probabilistic mixture-of-experts for efficient deep reinforcement learning","work_id":"11062be1-263d-496d-b7ce-e68bae5db606","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Hu- manoidbench: Simulated humanoid benchmark for whole-body locomotion and manipulation","work_id":"8e42d104-ddc6-4291-b477-f09257aa5c80","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"ST-MoE: Designing Stable and Transferable Sparse Expert Models","work_id":"b7581741-3f43-4528-a7d0-3af9e51a4d9f","ref_index":5,"cited_arxiv_id":"2202.08906","is_internal_anchor":true}],"resolved_work":5,"snapshot_sha256":"ee28e9b29883c85b4f1826a04aad8c39bb9715abda9f9e2c4e8a01cbb860e399","internal_anchors":2},"formal_canon":{"evidence_count":2,"snapshot_sha256":"4dc6796665eb157d9330c13b4cd04b9fb1af803a7c4231643e18ab5d192eff35"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}