{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "peng-abbeel-levine-van-de-panne-deepmimic-motion-imitation-2018", "title": "DeepMimic: learning simulated movement from example clips", "summary": "A simulated character can reach a target while moving in an implausible way. DeepMimic combines a task reward with a reward for matching a reference motion, so the policy learns both what to accomplish and how to move. Random starting points within a motion and early termination of unsuccessful attempts help training reach difficult parts of a skill. The work demonstrates dynamic movements and several ways to combine skills. Its contribution is this practical imitation-and-task learning approach, using existing reinforcement-learning machinery rather than introducing PPO itself.", "boundary": "The demonstrations are manually prepared short motion clips, with retargeting and control settings. The reported characters are simulated; an Atlas-shaped model is not deployment on a real Atlas robot. Training can take days per skill, and larger motion libraries or arbitrary transitions are not demonstrated. Phase-based motion limits retiming, finite perturbation tests do not guarantee robustness, and normalized rewards are not necessarily success probabilities or physically attainable maxima.", "area": "Physics-based character animation", "topic": "Artificial intelligence", "year": 2018, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "peng-abbeel-levine-van-de-panne-deepmimic-motion-imitation-2018", "target": { "record_id": "urn:uuid:1458a33a-c81f-42fc-af66-3870d573a39f", "version": 1, "content_hash": "sha256:3f9078b716e9fbd8ec5dfb0ccd822975733d4448fa427945be013ac50046a82a" }, "humans": [ "Xue Bin Peng", "Pieter Abbeel", "Sergey Levine", "Michiel van de Panne" ], "authors_complete": true, "human_role_note": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.", "agents": [], "related": [ "ppo-2017" ], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261017-arxiv-1804-02717v1", "title": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "url": "https://arxiv.org/abs/1804.02717v1", "date": "2018-04-08" } ], "points": "180000", "published_grade": { "id": "peng-abbeel-levine-van-de-panne-deepmimic-motion-imitation-2018", "entry_id": "peng-abbeel-levine-van-de-panne-deepmimic-motion-imitation-2018", "band": "A", "modifier": "+", "target": { "record_id": "urn:uuid:1458a33a-c81f-42fc-af66-3870d573a39f", "version": 1, "content_hash": "sha256:3f9078b716e9fbd8ec5dfb0ccd822975733d4448fa427945be013ac50046a82a" }, "what_changed": "A simulated character can reach a target while moving in an implausible way. DeepMimic combines a task reward with a reward for matching a reference motion, so the policy learns both what to accomplish and how to move. Random starting points within a motion and early termination of unsuccessful attempts help training reach difficult parts of a skill. The work demonstrates dynamic movements and several ways to combine skills. Its contribution is this practical imitation-and-task learning approach, using existing reinforcement-learning machinery rather than introducing PPO itself.", "scope": "A simulated character can reach a target while moving in an implausible way. DeepMimic combines a task reward with a reward for matching a reference motion, so the policy learns both what to accomplish and how to move. Random starting points within a motion and early termination of unsuccessful attempts help training reach difficult parts of a skill. The work demonstrates dynamic movements and several ways to combine skills. Its contribution is this practical imitation-and-task learning approach, using existing reinforcement-learning machinery rather than introducing PPO itself. The demonstrations are manually prepared short motion clips, with retargeting and control settings. The reported characters are simulated; an Atlas-shaped model is not deployment on a real Atlas robot. Training can take days per skill, and larger motion libraries or arbitrary transitions are not demonstrated. Phase-based motion limits retiming, finite perturbation tests do not guarantee robustness, and normalized rewards are not necessarily success probabilities or physically attainable maxima.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:1458a33a-c81f-42fc-af66-3870d573a39f", "version": 1, "content_hash": "sha256:3f9078b716e9fbd8ec5dfb0ccd822975733d4448fa427945be013ac50046a82a" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-23T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Xue Bin Peng", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "url": "https://arxiv.org/abs/1804.02717v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Pieter Abbeel", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "url": "https://arxiv.org/abs/1804.02717v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Sergey Levine", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "url": "https://arxiv.org/abs/1804.02717v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Michiel van de Panne", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "url": "https://arxiv.org/abs/1804.02717v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "url": "https://arxiv.org/abs/1804.02717v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "The demonstrations are manually prepared short motion clips, with retargeting and control settings. The reported characters are simulated; an Atlas-shaped model is not deployment on a real Atlas robot. Training can take days per skill, and larger motion libraries or arbitrary transitions are not demonstrated. Phase-based motion limits retiming, finite perturbation tests do not guarantee robustness, and normalized rewards are not necessarily success probabilities or physically attainable maxima.", "domain": "Physics-based character animation" }, "sources": [ { "citation": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "url": "https://arxiv.org/abs/1804.02717v1" } ], "statement": "# DeepMimic: learning simulated movement from example clips\n\n## Reaching the goal is not the whole movement\n\nA game character might reach a finish line by crawling awkwardly, even when its designer wanted a natural run. A reward for the destination alone does not fully specify the desired movement. DeepMimic combines a task objective with imitation of a reference motion. The agent learns a control policy, a rule for choosing actions, that aims to accomplish the task while keeping its movement similar to the example.\n\n## Two signals guide one policy\n\nAn imitation reward compares aspects of the simulated body with the reference clip, including its changing pose and motion. A task reward encourages the actual objective. The reference phase indicates where the character is within the clip. The researchers also start training episodes at different points in a motion, so difficult later stages need not wait for the agent to master every earlier stage. Existing reinforcement-learning algorithms perform the optimization.\n\n## Try a deliberately simple reward\n\nUse a classroom mixture that is not the paper’s chosen weighting: seventy percent imitation and thirty percent task reward. One candidate movement has imitation score one and task score zero, giving a total of 0.7. Another has imitation score 0.8 and task score one, giving 0.56 plus 0.3, or 0.86. The second wins this toy comparison. Change the weights and explain what the scoring now encourages; neither number is a measured probability of success.\n\n## The simulated character still needs preparation\n\nReference clips are short and prepared by people, with retargeting and controller settings. Training can take days for a skill. Some ways of combining clips may ignore less compatible examples, and fixed reference timing limits how freely a motion can be retimed. The paper shows simulated characters, including robot-shaped ones. It does not thereby establish that the policy ran on the corresponding physical robot or stayed safe under every disturbance.\n\n## A useful first assignment\n\nAsk your agent to build a small table of imitation scores, task scores and two alternative weight choices. Include a movement that reaches the goal in an unwanted way and a stylish movement that never finishes. Require an explanation of the trade-off before calculating a winner. Then locate one actual failure or limitation in the source. Keep this reward-design exercise separate from reproducing DeepMimic’s training or claiming that a toy score measures human-level movement.\n\n## Scientific scope\n\nThe demonstrations are manually prepared short motion clips, with retargeting and control settings. The reported characters are simulated; an Atlas-shaped model is not deployment on a real Atlas robot. Training can take days per skill, and larger motion libraries or arbitrary transitions are not demonstrated. Phase-based motion limits retiming, finite perturbation tests do not guarantee robustness, and normalized rewards are not necessarily success probabilities or physically attainable maxima.\n\n## Sources for this explanation\n\n- [DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills](https://arxiv.org/abs/1804.02717v1): Original 2018 method and experiments: imitation/task rewards, reference-state initialization, early termination, skill combinations, preparation and training limits.\n", "summary": "A simulated character can reach a target while moving in an implausible way. DeepMimic combines a task reward with a reward for matching a reference motion, so the policy learns both what to accomplish and how to move. Random starting points within a motion and early termination of unsuccessful attempts help training reach difficult parts of a skill. The work demonstrates dynamic movements and several ways to combine skills. Its contribution is this practical imitation-and-task learning approach, using existing reinforcement-learning machinery rather than introducing PPO itself.", "title": "DeepMimic: learning simulated movement from example clips" }, "record_id": "urn:uuid:1458a33a-c81f-42fc-af66-3870d573a39f", "version": 1, "source_ref": { "record_id": "urn:uuid:1458a33a-c81f-42fc-af66-3870d573a39f", "version": 1, "content_hash": "sha256:3f9078b716e9fbd8ec5dfb0ccd822975733d4448fa427945be013ac50046a82a" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:a3d59b222ced32a5987288e9bdcc901e684eb88bea4f135494c9dd76b42d0f7b" }, "target": { "record_id": "urn:uuid:1458a33a-c81f-42fc-af66-3870d573a39f", "version": 1, "content_hash": "sha256:3f9078b716e9fbd8ec5dfb0ccd822975733d4448fa427945be013ac50046a82a" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "peng-abbeel-levine-van-de-panne-deepmimic-motion-imitation-2018", "entry_id": "peng-abbeel-levine-van-de-panne-deepmimic-motion-imitation-2018", "revision": 1, "title": "DeepMimic: learning simulated movement from example clips", "context": "Physics-based character animation", "track": "Reusable scientific method, experimental result or historical framework", "band": "A", "modifier": "+", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-23", "evidence_cutoff": "2026-10-23", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "A simulated character can reach a target while moving in an implausible way. DeepMimic combines a task reward with a reward for matching a reference motion, so the policy learns both what to accomplish and how to move. Random starting points within a motion and early termination of unsuccessful attempts help training reach difficult parts of a skill. The work demonstrates dynamic movements and several ways to combine skills. Its contribution is this practical imitation-and-task learning approach, using existing reinforcement-learning machinery rather than introducing PPO itself.", "scope": "A simulated character can reach a target while moving in an implausible way. DeepMimic combines a task reward with a reward for matching a reference motion, so the policy learns both what to accomplish and how to move. Random starting points within a motion and early termination of unsuccessful attempts help training reach difficult parts of a skill. The work demonstrates dynamic movements and several ways to combine skills. Its contribution is this practical imitation-and-task learning approach, using existing reinforcement-learning machinery rather than introducing PPO itself. The demonstrations are manually prepared short motion clips, with retargeting and control settings. The reported characters are simulated; an Atlas-shaped model is not deployment on a real Atlas robot. Training can take days per skill, and larger motion libraries or arbitrary transitions are not demonstrated. Phase-based motion limits retiming, finite perturbation tests do not guarantee robustness, and normalized rewards are not necessarily success probabilities or physically attainable maxima.", "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The demonstrations are manually prepared short motion clips, with retargeting and control settings. The reported characters are simulated; an Atlas-shaped model is not deployment on a real Atlas robot. Training can take days per skill, and larger motion libraries or arbitrary transitions are not demonstrated. Phase-based motion limits retiming, finite perturbation tests do not guarantee robustness, and normalized rewards are not necessarily success probabilities or physically attainable maxima.", "references": [ { "title": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "url": "https://arxiv.org/abs/1804.02717v1", "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained." } ], "roles": [ { "who": "The named authors in the research record", "credit": "SCIENCE", "did": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred." } ], "target": { "record_id": "urn:uuid:1458a33a-c81f-42fc-af66-3870d573a39f", "version": 1, "content_hash": "sha256:3f9078b716e9fbd8ec5dfb0ccd822975733d4448fa427945be013ac50046a82a" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "180000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": { "id": "peng-abbeel-levine-van-de-panne-deepmimic-motion-imitation-2018", "target": { "record_id": "urn:uuid:1458a33a-c81f-42fc-af66-3870d573a39f", "version": 1, "content_hash": "sha256:3f9078b716e9fbd8ec5dfb0ccd822975733d4448fa427945be013ac50046a82a" }, "sections": [ { "title": "Reaching the goal is not the whole movement", "text": "A game character might reach a finish line by crawling awkwardly, even when its designer wanted a natural run. A reward for the destination alone does not fully specify the desired movement. DeepMimic combines a task objective with imitation of a reference motion. The agent learns a control policy, a rule for choosing actions, that aims to accomplish the task while keeping its movement similar to the example." }, { "title": "Two signals guide one policy", "text": "An imitation reward compares aspects of the simulated body with the reference clip, including its changing pose and motion. A task reward encourages the actual objective. The reference phase indicates where the character is within the clip. The researchers also start training episodes at different points in a motion, so difficult later stages need not wait for the agent to master every earlier stage. Existing reinforcement-learning algorithms perform the optimization." }, { "title": "Try a deliberately simple reward", "text": "Use a classroom mixture that is not the paper’s chosen weighting: seventy percent imitation and thirty percent task reward. One candidate movement has imitation score one and task score zero, giving a total of 0.7. Another has imitation score 0.8 and task score one, giving 0.56 plus 0.3, or 0.86. The second wins this toy comparison. Change the weights and explain what the scoring now encourages; neither number is a measured probability of success." }, { "title": "The simulated character still needs preparation", "text": "Reference clips are short and prepared by people, with retargeting and controller settings. Training can take days for a skill. Some ways of combining clips may ignore less compatible examples, and fixed reference timing limits how freely a motion can be retimed. The paper shows simulated characters, including robot-shaped ones. It does not thereby establish that the policy ran on the corresponding physical robot or stayed safe under every disturbance." }, { "title": "A useful first assignment", "text": "Ask your agent to build a small table of imitation scores, task scores and two alternative weight choices. Include a movement that reaches the goal in an unwanted way and a stylish movement that never finishes. Require an explanation of the trade-off before calculating a winner. Then locate one actual failure or limitation in the source. Keep this reward-design exercise separate from reproducing DeepMimic’s training or claiming that a toy score measures human-level movement." } ], "references": [ { "title": "DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills", "url": "https://arxiv.org/abs/1804.02717v1", "locator": "Original 2018 method and experiments: imitation/task rewards, reference-state initialization, early termination, skill combinations, preparation and training limits." } ] }, "reading_status": "EXPLANATORY_GUIDE" }