{"work":{"id":"38e3ca94-96f0-4b19-a355-0754931af8be","openalex_id":"https://openalex.org/W1191599655","doi":"10.48550/arxiv.1506.02438","arxiv_id":"1506.02438","raw_key":null,"title":"High-Dimensional Continuous Control Using Generalized Advantage Estimation","authors":null,"authors_text":"John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, Pieter Abbeel","year":2015,"venue":"cs.LG","abstract":"Policy gradient methods are an appealing approach in reinforcement learning because they directly optimize the cumulative reward and can straightforwardly be used with nonlinear function approximators such as neural networks. The two main challenges are the large number of samples typically required, and the difficulty of obtaining stable and steady improvement despite the nonstationarity of the incoming data. We address the first challenge by using value functions to substantially reduce the variance of policy gradient estimates at the cost of some bias, with an exponentially-weighted estimator of the advantage function that is analogous to TD(lambda). We address the second challenge by using trust region optimization procedure for both the policy and the value function, which are represented by neural networks.\n  Our approach yields strong empirical results on highly challenging 3D locomotion tasks, learning running gaits for bipedal and quadrupedal simulated robots, and learning a policy for getting the biped to stand up from starting out lying on the ground. In contrast to a body of prior work that uses hand-crafted policy representations, our neural network policies map directly from raw kinematics to joint torques. Our algorithm is fully model-free, and the amount of simulated experience required for the learning tasks on 3D bipeds corresponds to 1-2 weeks of real time.","external_url":"https://arxiv.org/abs/1506.02438","cited_by_count":1750,"metadata_source":"pith","metadata_fetched_at":"2026-08-05T02:28:24.338817+00:00","pith_arxiv_id":"1506.02438","created_at":"2026-05-09T05:45:23.324599+00:00","updated_at":"2026-08-05T02:28:24.338817+00:00","title_quality_ok":true,"display_title":"High-Dimensional Continuous Control Using Generalized Advantage Estimation","render_title":"High-Dimensional Continuous Control Using Generalized Advantage Estimation"},"hub":{"state":{"work_id":"38e3ca94-96f0-4b19-a355-0754931af8be","tier":"super_hub","tier_reason":"100+ Pith inbound or 10,000+ external citations","pith_inbound_count":193,"external_cited_by_count":1750,"distinct_field_count":25,"first_pith_cited_at":"2016-06-21T13:37:05+00:00","last_pith_cited_at":"2026-07-07T09:27:30+00:00","author_build_status":"needed","summary_status":"needed","contexts_status":"needed","graph_status":"needed","ask_index_status":"needed","reader_status":"not_needed","recognition_status":"not_needed","updated_at":"2026-08-21T10:49:20.960304+00:00","tier_text":"super_hub"},"tier":"super_hub","role_counts":[{"context_role":"background","n":18},{"context_role":"method","n":13},{"context_role":"baseline","n":2},{"context_role":"dataset","n":1}],"polarity_counts":[{"context_polarity":"background","n":16},{"context_polarity":"use_method","n":13},{"context_polarity":"baseline","n":3},{"context_polarity":"unclear","n":2}],"runs":{"ask_index":{"job_type":"ask_index","status":"succeeded","result":{"title":"High-Dimensional Continuous Control Using Generalized Advantage Estimation","claims":[{"claim_text":"Policy gradient methods are an appealing approach in reinforcement learning because they directly optimize the cumulative reward and can straightforwardly be used with nonlinear function approximators such as neural networks. The two main challenges are the large number of samples typically required, and the difficulty of obtaining stable and steady improvement despite the nonstationarity of the incoming data. We address the first challenge by using value functions to substantially reduce the variance of policy gradient estimates at the cost of some bias, with an exponentially-weighted estimat","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"L(θ, θk) =E s,a∼πθk \" πθ(a|s) πθk(a|s) Aπθk (s, a) # ,(7) and ¯DKL(θ||θ k)is the average KL-divergence between policies, both computed using states visited and corresponding actions sampled by the old policy. In practice, the authors solve an approximate optimization problem derived from a Taylor expansion of Eq. (6). 3) Proximal Policy Optimization (PPO) [44]:emerged as an alternative that retains the stability of TRPO but with a simpler, first-order optimization approach. PPO's key insight is ","claim_type":"method","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"The stochastic policy is then πθ(ai t |s i t, µt) =N ai t ¯µi t,diag(σ 2) \u0001 ,(25) so the forward pass executes only thekselected experts and the small router. e) Actor objective with MoE terms:Under MAPPO, the actor minimizes Lactor =E[L PPO(θ)]−λ ent E \u0002 H πθ(· |s i t, µt) \u0001\u0003 ,(26) withλ ent ≥0. The critic and advantage estimator follow standard PPO and GAE [45], [46] under CTDE [47]. f) Why MoE here:Top-Ktoken choice routing scales capacity while keeping per-step cost near that of a small dens","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"the policy loss: Lπ =E t [min (ζt(θ)·A t,clip(ζ t(θ),1−ϵ,1 +ϵ)·A t)] −β H(π(s t)), and the value loss:L v =A 2 t . The advantage functionA t quantifies the relative quality of an action compared to the expected behavior under the current policy. To computeA t more accurately and robustly, we employ the Generalized Advantage Estimation (GAE) technique [22]. The termζ t(θ) is the likelihood ratio between the new and old policies. The parameterϵdefines the clipping range that restricts the magnitud","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Let πθ(a|s) be the policy parameterized byθ, and πθold(a|s) be the old policy from the previous iteration. The surrogate objective function for PPO is defined as: LCLIP (θ) = ˆEt h min \u0010 rt(θ) ˆAt,clip(rt(θ), 1 − ϵ, 1 + ϵ) ˆAt \u0011i (2) where rt(θ) = πθ(at|st) πθold (at|st) is the probability ratio, ˆAt is the estimated advantage at time stept, and ϵ is a hyperparameter that controls the clipping range. Generalized Advantage Estimation [20] is a technique used to estimate the advantage function mor","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"is_equiv() function provided in the official github repository of [17]. All evaluations use temperature 0.5, Top-p=0.95, and max response length of2048tokens. Random seed is42across all experiments. Baselines.We compare our trained controller with four pruning baselines: frequency-based selection, recon- struction loss minimization [24], random selection, and Wanda (structured) [34]. For all baselines, we use 128 prompts randomly drawn from the Nemotron Post-training Dataset v2 as the calibratio","claim_type":"baseline","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"2017. arXiv:1707.06347 [cs], 10.48550/arXiv.1707.06347. [38] N. Hammami and K. K. Nguyen, \"On-Policy vs. Off-Policy Deep Reinforcement Learning for Resource Allocation in Open Radio Access Network,\"2022 IEEE Wireless Communica- tions and Networking Conference (WCNC), Austin, TX, USA, IEEE, Apr. 2022, pp. 1461-1466, 10.1109/WCNC51071.2022.9771605. [39] J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel, \"High-Dimensional Contin- uous Control Using Generalized Advantage Estimation,\" Oct. ","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks High-Dimensional Continuous Control Using Generalized Advantage Estimation because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (13 contexts).","role_counts":[{"n":13,"context_role":"background"},{"n":13,"context_role":"method"},{"n":2,"context_role":"baseline"},{"n":1,"context_role":"dataset"}]},"error":null,"updated_at":"2026-05-19T14:01:28.552380+00:00"},"author_expand":{"job_type":"author_expand","status":"succeeded","result":{"authors_linked":[{"id":"298fbbc8-0497-4319-a5d8-d4bcef3f7f3d","orcid":null,"display_name":"John Schulman"},{"id":"237f757e-89b2-4ae7-8e57-50d5f1cc4a37","orcid":null,"display_name":"Philipp Moritz"},{"id":"ee56e6c3-f424-4a4b-852d-8ab2deb6dc65","orcid":null,"display_name":"Sergey Levine"},{"id":"988f99d5-de7a-4b09-af65-c524d652d612","orcid":null,"display_name":"Michael Jordan"},{"id":"0f251250-6471-4035-9a96-4e2bd1ce0707","orcid":null,"display_name":"Pieter Abbeel"}]},"error":null,"updated_at":"2026-05-19T14:01:28.988301+00:00"},"context_extract":{"job_type":"context_extract","status":"succeeded","result":{"enqueued_papers":25},"error":null,"updated_at":"2026-05-14T09:18:16.073395+00:00"},"graph_features":{"job_type":"graph_features","status":"succeeded","result":{"co_cited":[{"title":"Proximal Policy Optimization Algorithms","work_id":"240c67fe-d14d-4520-91c1-38a4e272ca19","shared_citers":54},{"title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","work_id":"c5006563-f3ec-438a-9e35-b7b484f34828","shared_citers":22},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":12},{"title":"Adam: A Method for Stochastic Optimization","work_id":"1910796d-9b52-4683-bf5c-de9632c1028b","shared_citers":10},{"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","shared_citers":10},{"title":"DAPO: An Open-Source LLM Reinforcement Learning System at Scale","work_id":"64019d00-0b11-4bbd-b173-b46c8fad0157","shared_citers":9},{"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","shared_citers":8},{"title":"Group Sequence Policy Optimization","work_id":"3a98b53b-9f52-4d95-adf7-89353c0a9a65","shared_citers":6},{"title":"Measuring Mathematical Problem Solving With the MATH Dataset","work_id":"50652ac6-fb7c-4675-a2c2-159c241feb17","shared_citers":6},{"title":"Understanding R1-Zero-Like Training: A Critical Perspective","work_id":"ec354f3b-9484-4a0c-94c8-92d4d0260835","shared_citers":6},{"title":"OpenAI o1 System Card","work_id":"68d3c334-0fc9-49e3-b7b0-a69afae933e2","shared_citers":5},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":5},{"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","shared_citers":5},{"title":"Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine","work_id":"a0c5b73d-c242-4cb2-b4d4-fbffb504fe33","shared_citers":5},{"title":"Andrychowicz, A","work_id":"67073b84-d283-4a63-ba53-c9e65061ec57","shared_citers":4},{"title":"Dota 2 with Large Scale Deep Reinforcement Learning","work_id":"b047dc18-e9a3-4d11-8ff6-cd59d41a6357","shared_citers":4},{"title":"HybridFlow: A Flexible and Efficient RLHF Framework","work_id":"7eb9c9f4-b322-4bba-8011-09ff8d6ad801","shared_citers":4},{"title":"Kimi k1.5: Scaling Reinforcement Learning with LLMs","work_id":"bff96ab1-bd6a-4585-be23-74fdb51969c7","shared_citers":4},{"title":"Playing Atari with Deep Reinforcement Learning","work_id":"736a8ddf-e365-4940-ad58-4699fddedb86","shared_citers":4},{"title":"Qwen2.5 Technical Report","work_id":"d8432992-4980-4a81-85c7-9fa2c2b87f85","shared_citers":4},{"title":"REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalization","work_id":"557f9e99-cb00-4dd2-92fd-67ddcddbb35d","shared_citers":4},{"title":"Training Diffusion Models with Reinforcement Learning","work_id":"67684dda-3930-452a-b91a-36cbb8e2e219","shared_citers":4},{"title":"$\\pi_0$: A Vision-Language-Action Flow Model for General Robot Control","work_id":"f790abdc-a796-482f-a40d-f8ee035ecfc2","shared_citers":3},{"title":"Deep reinforcement learning and the deadly triad","work_id":"de214ead-4cb0-4abd-be3d-ae3389f55e9b","shared_citers":3}],"time_series":[{"n":1,"year":2016},{"n":1,"year":2019},{"n":1,"year":2024},{"n":2,"year":2025},{"n":61,"year":2026}],"dependency_candidates":[]},"error":null,"updated_at":"2026-05-14T09:18:14.137206+00:00"},"identity_refresh":{"job_type":"identity_refresh","status":"succeeded","result":{"items":[{"title":"Qwen3 Technical Report","outcome":"unchanged","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","resolver":"local_arxiv","confidence":0.98,"old_work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e"}],"counts":{"fixed":0,"merged":0,"unchanged":1,"quarantined":0,"needs_external_resolution":0},"errors":[],"attempted":1},"error":null,"updated_at":"2026-05-14T09:18:10.182535+00:00"},"role_polarity":{"job_type":"role_polarity","status":"succeeded","result":{"title":"High-Dimensional Continuous Control Using Generalized Advantage Estimation","claims":[{"claim_text":"Policy gradient methods are an appealing approach in reinforcement learning because they directly optimize the cumulative reward and can straightforwardly be used with nonlinear function approximators such as neural networks. The two main challenges are the large number of samples typically required, and the difficulty of obtaining stable and steady improvement despite the nonstationarity of the incoming data. We address the first challenge by using value functions to substantially reduce the variance of policy gradient estimates at the cost of some bias, with an exponentially-weighted estimat","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"L(θ, θk) =E s,a∼πθk \" πθ(a|s) πθk(a|s) Aπθk (s, a) # ,(7) and ¯DKL(θ||θ k)is the average KL-divergence between policies, both computed using states visited and corresponding actions sampled by the old policy. In practice, the authors solve an approximate optimization problem derived from a Taylor expansion of Eq. (6). 3) Proximal Policy Optimization (PPO) [44]:emerged as an alternative that retains the stability of TRPO but with a simpler, first-order optimization approach. PPO's key insight is ","claim_type":"method","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"The stochastic policy is then πθ(ai t |s i t, µt) =N ai t ¯µi t,diag(σ 2) \u0001 ,(25) so the forward pass executes only thekselected experts and the small router. e) Actor objective with MoE terms:Under MAPPO, the actor minimizes Lactor =E[L PPO(θ)]−λ ent E \u0002 H πθ(· |s i t, µt) \u0001\u0003 ,(26) withλ ent ≥0. The critic and advantage estimator follow standard PPO and GAE [45], [46] under CTDE [47]. f) Why MoE here:Top-Ktoken choice routing scales capacity while keeping per-step cost near that of a small dens","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"the policy loss: Lπ =E t [min (ζt(θ)·A t,clip(ζ t(θ),1−ϵ,1 +ϵ)·A t)] −β H(π(s t)), and the value loss:L v =A 2 t . The advantage functionA t quantifies the relative quality of an action compared to the expected behavior under the current policy. To computeA t more accurately and robustly, we employ the Generalized Advantage Estimation (GAE) technique [22]. The termζ t(θ) is the likelihood ratio between the new and old policies. The parameterϵdefines the clipping range that restricts the magnitud","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Let πθ(a|s) be the policy parameterized byθ, and πθold(a|s) be the old policy from the previous iteration. The surrogate objective function for PPO is defined as: LCLIP (θ) = ˆEt h min \u0010 rt(θ) ˆAt,clip(rt(θ), 1 − ϵ, 1 + ϵ) ˆAt \u0011i (2) where rt(θ) = πθ(at|st) πθold (at|st) is the probability ratio, ˆAt is the estimated advantage at time stept, and ϵ is a hyperparameter that controls the clipping range. Generalized Advantage Estimation [20] is a technique used to estimate the advantage function mor","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"is_equiv() function provided in the official github repository of [17]. All evaluations use temperature 0.5, Top-p=0.95, and max response length of2048tokens. Random seed is42across all experiments. Baselines.We compare our trained controller with four pruning baselines: frequency-based selection, recon- struction loss minimization [24], random selection, and Wanda (structured) [34]. For all baselines, we use 128 prompts randomly drawn from the Nemotron Post-training Dataset v2 as the calibratio","claim_type":"baseline","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"2017. arXiv:1707.06347 [cs], 10.48550/arXiv.1707.06347. [38] N. Hammami and K. K. Nguyen, \"On-Policy vs. Off-Policy Deep Reinforcement Learning for Resource Allocation in Open Radio Access Network,\"2022 IEEE Wireless Communica- tions and Networking Conference (WCNC), Austin, TX, USA, IEEE, Apr. 2022, pp. 1461-1466, 10.1109/WCNC51071.2022.9771605. [39] J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel, \"High-Dimensional Contin- uous Control Using Generalized Advantage Estimation,\" Oct. ","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks High-Dimensional Continuous Control Using Generalized Advantage Estimation because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (13 contexts).","role_counts":[{"n":13,"context_role":"background"},{"n":13,"context_role":"method"},{"n":2,"context_role":"baseline"},{"n":1,"context_role":"dataset"}]},"error":null,"updated_at":"2026-05-19T14:01:28.991683+00:00"},"summary_claims":{"job_type":"summary_claims","status":"succeeded","result":{"title":"High-Dimensional Continuous Control Using Generalized Advantage Estimation","claims":[{"claim_text":"Policy gradient methods are an appealing approach in reinforcement learning because they directly optimize the cumulative reward and can straightforwardly be used with nonlinear function approximators such as neural networks. The two main challenges are the large number of samples typically required, and the difficulty of obtaining stable and steady improvement despite the nonstationarity of the incoming data. We address the first challenge by using value functions to substantially reduce the variance of policy gradient estimates at the cost of some bias, with an exponentially-weighted estimat","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks High-Dimensional Continuous Control Using Generalized Advantage Estimation because it crossed a citation-hub threshold.","role_counts":[]},"error":null,"updated_at":"2026-05-14T09:18:07.913094+00:00"}},"summary":{"title":"High-Dimensional Continuous Control Using Generalized Advantage Estimation","claims":[{"claim_text":"Policy gradient methods are an appealing approach in reinforcement learning because they directly optimize the cumulative reward and can straightforwardly be used with nonlinear function approximators such as neural networks. The two main challenges are the large number of samples typically required, and the difficulty of obtaining stable and steady improvement despite the nonstationarity of the incoming data. We address the first challenge by using value functions to substantially reduce the variance of policy gradient estimates at the cost of some bias, with an exponentially-weighted estimat","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks High-Dimensional Continuous Control Using Generalized Advantage Estimation because it crossed a citation-hub threshold.","role_counts":[]},"graph":{"co_cited":[{"title":"Proximal Policy Optimization Algorithms","work_id":"240c67fe-d14d-4520-91c1-38a4e272ca19","shared_citers":54},{"title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","work_id":"c5006563-f3ec-438a-9e35-b7b484f34828","shared_citers":22},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":12},{"title":"Adam: A Method for Stochastic Optimization","work_id":"1910796d-9b52-4683-bf5c-de9632c1028b","shared_citers":10},{"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","shared_citers":10},{"title":"DAPO: An Open-Source LLM Reinforcement Learning System at Scale","work_id":"64019d00-0b11-4bbd-b173-b46c8fad0157","shared_citers":9},{"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","shared_citers":8},{"title":"Group Sequence Policy Optimization","work_id":"3a98b53b-9f52-4d95-adf7-89353c0a9a65","shared_citers":6},{"title":"Measuring Mathematical Problem Solving With the MATH Dataset","work_id":"50652ac6-fb7c-4675-a2c2-159c241feb17","shared_citers":6},{"title":"Understanding R1-Zero-Like Training: A Critical Perspective","work_id":"ec354f3b-9484-4a0c-94c8-92d4d0260835","shared_citers":6},{"title":"OpenAI o1 System Card","work_id":"68d3c334-0fc9-49e3-b7b0-a69afae933e2","shared_citers":5},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":5},{"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","shared_citers":5},{"title":"Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine","work_id":"a0c5b73d-c242-4cb2-b4d4-fbffb504fe33","shared_citers":5},{"title":"Andrychowicz, A","work_id":"67073b84-d283-4a63-ba53-c9e65061ec57","shared_citers":4},{"title":"Dota 2 with Large Scale Deep Reinforcement Learning","work_id":"b047dc18-e9a3-4d11-8ff6-cd59d41a6357","shared_citers":4},{"title":"HybridFlow: A Flexible and Efficient RLHF Framework","work_id":"7eb9c9f4-b322-4bba-8011-09ff8d6ad801","shared_citers":4},{"title":"Kimi k1.5: Scaling Reinforcement Learning with LLMs","work_id":"bff96ab1-bd6a-4585-be23-74fdb51969c7","shared_citers":4},{"title":"Playing Atari with Deep Reinforcement Learning","work_id":"736a8ddf-e365-4940-ad58-4699fddedb86","shared_citers":4},{"title":"Qwen2.5 Technical Report","work_id":"d8432992-4980-4a81-85c7-9fa2c2b87f85","shared_citers":4},{"title":"REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalization","work_id":"557f9e99-cb00-4dd2-92fd-67ddcddbb35d","shared_citers":4},{"title":"Training Diffusion Models with Reinforcement Learning","work_id":"67684dda-3930-452a-b91a-36cbb8e2e219","shared_citers":4},{"title":"$\\pi_0$: A Vision-Language-Action Flow Model for General Robot Control","work_id":"f790abdc-a796-482f-a40d-f8ee035ecfc2","shared_citers":3},{"title":"Deep reinforcement learning and the deadly triad","work_id":"de214ead-4cb0-4abd-be3d-ae3389f55e9b","shared_citers":3}],"time_series":[{"n":1,"year":2016},{"n":1,"year":2019},{"n":1,"year":2024},{"n":2,"year":2025},{"n":61,"year":2026}],"dependency_candidates":[]},"authors":[{"id":"298fbbc8-0497-4319-a5d8-d4bcef3f7f3d","orcid":null,"display_name":"John Schulman","source":"manual","import_confidence":0.72},{"id":"988f99d5-de7a-4b09-af65-c524d652d612","orcid":null,"display_name":"Michael Jordan","source":"manual","import_confidence":0.72},{"id":"237f757e-89b2-4ae7-8e57-50d5f1cc4a37","orcid":null,"display_name":"Philipp Moritz","source":"manual","import_confidence":0.72},{"id":"0f251250-6471-4035-9a96-4e2bd1ce0707","orcid":null,"display_name":"Pieter Abbeel","source":"manual","import_confidence":0.72},{"id":"ee56e6c3-f424-4a4b-852d-8ab2deb6dc65","orcid":null,"display_name":"Sergey Levine","source":"manual","import_confidence":0.72}]}}