{"paper":{"title":"TD-MPC2: Scalable, Robust World Models for Continuous Control","license":"http://creativecommons.org/licenses/by/4.0/","headline":"TD-MPC2 achieves significantly better performance than baselines on 104 continuous control tasks using one fixed set of hyperparameters.","cross_cats":["cs.AI","cs.CV","cs.RO"],"primary_cat":"cs.LG","authors_text":"Hao Su, Nicklas Hansen, Xiaolong Wang","submitted_at":"2023-10-25T17:57:07Z","abstract_excerpt":"TD-MPC is a model-based reinforcement learning (RL) algorithm that performs local trajectory optimization in the latent space of a learned implicit (decoder-free) world model. In this work, we present TD-MPC2: a series of improvements upon the TD-MPC algorithm. We demonstrate that TD-MPC2 improves significantly over baselines across 104 online RL tasks spanning 4 diverse task domains, achieving consistently strong results with a single set of hyperparameters. We further show that agent capabilities increase with model and data size, and successfully train a single 317M parameter agent to perfo"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We demonstrate that TD-MPC2 improves significantly over baselines across 104 online RL tasks spanning 4 diverse task domains, achieving consistently strong results with a single set of hyperparameters. We further show that agent capabilities increase with model and data size, and successfully train a single 317M parameter agent to perform 80 tasks across multiple task domains, embodiments, and action spaces.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The reported gains rely on the assumption that the chosen 104 tasks and four domains are representative enough that a single hyperparameter set will continue to work when the method is applied to new, unseen continuous-control problems.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"TD-MPC2 scales an implicit world-model RL method to a 317M-parameter agent that masters 80 tasks across four domains with a single hyperparameter configuration.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"TD-MPC2 achieves significantly better performance than baselines on 104 continuous control tasks using one fixed set of hyperparameters.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"9d9457d575688ec95175cba1bfd515c0b36a0c0b76d352ed9c39c36f271ca75f"},"source":{"id":"2310.16828","kind":"arxiv","version":2},"verdict":{"id":"199b3dec-94d2-402a-9f97-e5c6f06258a3","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T17:21:09.097426Z","strongest_claim":"We demonstrate that TD-MPC2 improves significantly over baselines across 104 online RL tasks spanning 4 diverse task domains, achieving consistently strong results with a single set of hyperparameters. We further show that agent capabilities increase with model and data size, and successfully train a single 317M parameter agent to perform 80 tasks across multiple task domains, embodiments, and action spaces.","one_line_summary":"TD-MPC2 scales an implicit world-model RL method to a 317M-parameter agent that masters 80 tasks across four domains with a single hyperparameter configuration.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The reported gains rely on the assumption that the chosen 104 tasks and four domains are representative enough that a single hyperparameter set will continue to work when the method is applied to new, unseen continuous-control problems.","pith_extraction_headline":"TD-MPC2 achieves significantly better performance than baselines on 104 continuous control tasks using one fixed set of hyperparameters."},"references":{"count":162,"sample":[{"doi":"","year":2016,"title":"Layer normalization","work_id":"d34753e2-6f32-4f79-9705-ea5c146ebd8c","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Video pretraining (vpt): Learning to act by watching unlabeled online videos","work_id":"3ba8312e-03af-4ed5-8f67-041856cbe023","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"A distributional perspective on reinforcement learning","work_id":"b1fb027a-a8d8-4a70-bd30-47ca98dc5f84","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1957,"title":"A markovian decision process","work_id":"27354393-5813-4731-88c1-6372a303aefe","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1901,"title":"Language models are few-shot learners","work_id":"06905937-01c1-4635-ab7b-151e446701ad","ref_index":7,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":162,"snapshot_sha256":"9705adb397d177f7637798c09f0e4dec6fd0adebbbac081623d69202093275db","internal_anchors":12},"formal_canon":{"evidence_count":3,"snapshot_sha256":"5848add13880657af845b8fb0ab03a545150627f2fb09cf60d22024a54a666cf"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}