GSBR: Geometric Support Behavior Regularization for Online RL
Charles Xu, Jacob Berg, Liyiming Ke, Sergey Levine
AI 速览：提出 GSBR，用动作到基策略支撑集的距离做正则化，在支撑集内不惩罚、支撑集外平滑惩罚，以兼顾在线 RL 的探索改进与稳定性。
Regularizing online reinforcement learning (RL) toward a pretrained base policy can improve sample efficiency and learning stability, but regularization that is too restrictive limits how much the policy can improve. Base policies -- typically diffusion or flow-matching models trained by imitation learning on large, diverse datasets -- capture many behaviors that are reasonable but suboptimal. Existing regularizers using such base policies are too restrictive in one of two ways: behavioral cloning regularizers pull the actor toward the full base-policy distribution, penalizing it for discarding suboptimal modes, while hard support constraints forbid improvements that lie just outside the base policy's support. We propose Geometric Support Behavior Regularization (GSBR), which penalizes actions by their distance from the base policy's support. Actions within the support incur no penalty, so the actor can concentrate on high-value modes; outside the support, where value estimates are less reliable, the penalty grows smoothly with distance while still permitting improvement when the value gained outweighs the penalty. GSBR estimates this distance from a finite set of base-policy samples, so it requires only sample access to the base policy. Across 19 tasks on four robotic benchmarks, GSBR raises the average success rate from 33% for the base policy to 95%, achieving the highest average success rate and throughput among all compared methods, with about 23% higher normalized throughput than the next-best RL method. Project website and open-source code at https://charlesxu0124.github.io/gsbr/ .
Stabilizing On-Policy Targeted Guidance in Agentic Reinforcement Learning
Kevin Frans, Alexander M Rush, Sergey Levine, Pieter Abbeel, Ashvin Nair
AI 速览：作者提出并分析“定向引导”框架，通过同策略自蒸馏从逐轮文本反馈中学习，并在智能体多轮软件工程环境中探索如何稳定训练且不损害基线RL性能。
Naive on-policy RL is often bottlenecked by the capabilities of the base model --- insufficient coverage can result in an inability to learn from certain rewards. At the same time, rewards in modern post-training systems often come from rich textual sources, e.g. judge models with a rubric of criteria. Can we utilize such structures to provide a rich directional signal that enables improving on otherwise difficult to optimize rewards? In this work, we examine a framework we refer to as targeted guidance that learns from per-turn textual feedback via an on-policy self-distillation objective. Our focus is not only on directly improving hard-to-optimize behaviors, but on how to do so throughout the training process without degrading baseline RL performance. Over a range of settings across an agentic multi-turn software engineering environment, we find success by combining targeted self-distillation with a per-criteria RL reward, and by limiting feedback to turns where the policy specifically fails. We further examine the design choices that lead to stable entropy and reduce oscillatory behavior throughout training.
Jakub Grudzien Kuba, Benjamin Kurt Miller, Sergey Levine, Pieter Abbeel
Recent advances in deep learning inspired neural network-based approaches to computational materials discovery (CMD). A plethora of problems in this field involve finding materials that optimize a target property. Nevertheless, the increasingly popular generative modeling methods are ineffective at boldly exploring attractive regions of the materials space due to their maximum likelihood training. In this work, we offer an alternative CMD technique based on offline model-based optimization (MBO) that fuses direct optimization of a target material property into generation. To that end, we introduce a domain-specific model, dubbed CliqueFlowmer, that incorporates recent advances of clique-based MBO into transformer and flow generation. We validate this model's optimization abilities and show that materials it produces strongly outperform those from generative baselines. To support specialized materials discovery applications and broader interdisciplinary research, we release our code (https://github.com/znowu/CliqueFlowmer), model weights (https://huggingface.co/iamkuba/CliqueFlowmer), and an interactive demo (https://cliqueflowmer.znowu.workers.dev).
Chongyi Zheng, Zhiyuan Zhou, Qiyang Li, Benjamin Eysenbach, Sergey Levine
Learning accurate value functions is central to reinforcement learning (RL), yet remains challenging because supervision is limited and indirect: ground-truth values are unavailable, and each transition in the data only provides an immediate reward that can be used to construct bootstrapped targets as supervision. This challenge is particularly pronounced when data is limited or reward signals are sparse, as is common in real-world applications such as robotics. We argue that the key to accurate value learning in this limited data regime is extracting richer learning signals from the available data. In this paper, we introduce Value Forcing, a new approach that uses flow matching to estimate return distributions and enables accurate value learning from limited data. Our key insight is that we can extract richer learning signals by enforcing Bellman consistency at every noise level. Empirically, Value Forcing improves the accuracy of estimating the return distribution, and prevents the distribution from collapsing to a single value, as occurs with prior methods. We instantiate Value Forcing with two practical value-learning algorithms: VF, an actor-critic method, and IVF, an implicit value-learning method. We evaluate both algorithms across $26$ long-horizon tasks in the offline RL setting and find that they outperform prior value-learning methods, with particularly large gains when data is limited. Our project website is available at https://chongyi-zheng.github.io/value-forcing.
Qiyang Li, Zhiyuan Zhou, Charles Xu, Sergey Levine
Modern robot learning systems feed on high-quality demonstration data, and their performance hinges on coverage. Without good coverage, small errors compound and the policy can quickly go into out-of-distribution states from which it cannot recover. Yet there is no principled way of collecting high-coverage data, and data collection often relies on heuristic instructions to human operators. Purely online reinforcement learning (RL) explores at the cost of safety and sample efficiency, while DAgger-style interventions require the policy to fail by chance before a human can correct it. In this paper, we ask instead whether we can modify how the robot behaves during data collection to collect more useful data. We propose Data Inoculation (DI), a shared-control framework in which a human provide closed-loop teleoperation commands while a residual policy perturbs these commands. The residual policy is trained online, on the growing dataset, with the same reward-maximizing objective later used for policy extraction, so its perturbations steer data collection toward the deviations a policy trained on the data would make. These perturbations expose the robot to a wider range of states, while the human's continuous corrections provide recovery demonstrations, "inoculating" the dataset against the learner's future mistakes while the human remains in control. On three challenging real-world robotic tasks, DI consistently improves or matches the success rate of policies trained with both behavioral cloning and offline RL, showing DI as an effective data collection strategy for improving downstream policy performance.
Kai Nelson, Tobias Kreiman, Sergey Levine, Aditi S. Krishnapriyan
A fundamental challenge in science and engineering is the simulation-to-experiment gap. While we often possess prior knowledge of physical laws, these physical laws can be too difficult to solve exactly for complex systems. Such systems are commonly modeled using simulators, which impose computational approximations. Meanwhile, experimental measurements more faithfully represent the real world, but experimental data typically consists of observations that only partially reflect the system's full underlying state. We propose a data-driven distribution alignment framework that bridges this simulation-to-experiment gap by pre-training a generative model on fully observed (but imperfect) simulation data, then aligning it with partial (but real) observations of experimental data. While our method is domain-agnostic, we ground our approach in the physical sciences by introducing Adversarial Distribution Alignment (ADA). This method aligns a generative model of atomic positions -- initially trained on a simulated Boltzmann distribution -- with the distribution of experimental observations. We prove that our method recovers the target observable distribution, even with multiple, potentially correlated observables. We also empirically validate our framework on synthetic, molecular, and experimental protein data, demonstrating that it can align generative models with diverse observables. Our code is available at https://kaityrusnelson.com/ada/.
Zhiyuan Zhou, Andy Peng, Charles Xu, Qiyang Li, Jost Tobias Springenberg, Kevin Frans, Sergey Levine
Expressive continuous control policies, such as diffusion and flow models, form the backbone of recent advances in scaling imitation learning for simulated and real robot control. While they are known to scale stably in the supervised imitation learning setting, incorporating them into reinforcement learning (RL) pipelines for policy improvement has proven more difficult. It often requires specialized training objectives or backpropagating through denoising processes, which cause well-known issues with stability and affect scalability. In this paper we study the question of whether simple policy improvement schemes at test time alone, leaving stable supervised policy training intact, can be a competitive alternative which sidesteps these issues. To this end, we propose QGF (Q-Guided Flow), an RL algorithm that performs policy optimization entirely at test time. QGF works by pre-training both a reference flow policy (via a standard behavioral cloning objective) and a value function critic and, at test time, using the value gradient to guide the reference policy to generate higher-value actions without any additional policy learning. Empirically, QGF outperforms prior test-time RL methods on single-task and goal-conditioned offline RL benchmarks with high-dimensional action spaces, and is competitive with state-of-the-art training-time algorithms while being much cheaper to run. Moreover, it exhibits favorable scaling with model size by avoiding the instability of actor-critic training, offering a practical and effective alternative RL algorithm with expressive policies.
Transitive RL: Value Learning via Divide and Conquer
Seohong Park, Aditya Oberai, Pranav Atreya, Sergey Levine
AI 速览：提出 TRL，将目标条件强化学习中的三角不等式结构转化为分治式价值更新规则，用于离线 GCRL 的价值学习。
In this work, we present Transitive Reinforcement Learning (TRL), a new value learning algorithm based on a divide-and-conquer paradigm. TRL is designed for offline goal-conditioned reinforcement learning (GCRL) problems, where the aim is to find a policy that can reach any state from any other state in the smallest number of steps. TRL converts a triangle inequality structure present in GCRL into a practical divide-and-conquer value update rule. This has several advantages compared to alternative value learning paradigms. Compared to temporal difference (TD) methods, TRL suffers less from bias accumulation, as in principle it only requires $O(\log T)$ recursions (as opposed to $O(T)$ in TD learning) to handle a length-$T$ trajectory. Unlike Monte Carlo methods, TRL suffers less from high variance as it performs dynamic programming. Experimentally, we show that TRL achieves the best performance in highly challenging, long-horizon benchmark tasks compared to previous offline GCRL algorithms.
Seohong Park, Kevin Frans, Deepinder Mann, Benjamin Eysenbach, Aviral Kumar, Sergey Levine
AI 速览：该论文研究离线强化学习的可扩展性，发现长视野是主要瓶颈，并提出通过视野缩减技术（如SHARSA）提升性能与扩展性。
In this work, we study the scalability of offline reinforcement learning (RL) algorithms. In principle, a truly scalable offline RL algorithm should be able to solve any given problem, regardless of its complexity, given sufficient data, compute, and model capacity. We investigate if and how current offline RL algorithms match up to this promise on diverse, challenging, previously unsolved tasks, using datasets up to 1000x larger than typical offline RL datasets. We observe that despite scaling up data, many existing offline RL algorithms exhibit poor scaling behavior, saturating well below the maximum performance. We hypothesize that the horizon is the main cause behind the poor scaling of offline RL. We empirically verify this hypothesis through several analysis experiments, showing that long horizons indeed present a fundamental barrier to scaling up offline RL. We then show that various horizon reduction techniques substantially enhance scalability on challenging tasks. Based on our insights, we also introduce a minimal yet scalable method named SHARSA that effectively reduces the horizon. SHARSA achieves the best asymptotic performance and scaling behavior among our evaluation methods, showing that explicitly reducing the horizon unlocks the scalability of offline RL. Code: https://github.com/seohongpark/horizon-reduction
SWEET-RL: Training Multi-Turn LLM Agents on Collaborative Reasoning Tasks
Yifei Zhou, Song Jiang, Yuandong Tian, Jason Weston, Sergey Levine, Sainbayar Sukhbaatar, Xian Li
AI 速览：提出 ColBench 多轮人机协作基准与 SWEET-RL 算法，利用训练时额外信息训练评论家模型提供步级奖励，以改进多轮 LLM 智能体的信用分配。
Large language model (LLM) agents need to perform multi-turn interactions in real-world tasks. However, existing multi-turn RL algorithms for optimizing LLM agents fail to perform effective credit assignment over multiple turns while leveraging the generalization capabilities of LLMs and it remains unclear how to develop such algorithms. To study this, we first introduce a new benchmark, ColBench, where an LLM agent interacts with a human collaborator over multiple turns to solve realistic tasks in backend programming and frontend design. Building on this benchmark, we propose a novel RL algorithm, SWEET-RL (RL with Step-WisE Evaluation from Training-time information), that uses a carefully designed optimization objective to train a critic model with access to additional training-time information. The critic provides step-level rewards for improving the policy model. Our experiments demonstrate that SWEET-RL achieves a 6% absolute improvement in success and win rates on ColBench compared to other state-of-the-art multi-turn RL algorithms, enabling Llama-3.1-8B to match or exceed the performance of GPT4-o in realistic collaborative content creation.
Learning Temporal Distances: Contrastive Successor Features Can Provide a Metric Structure for Decision-Making
Vivek Myers, Chongyi Zheng, Anca Dragan, Sergey Levine, Benjamin Eysenbach
AI 速览：该论文提出通过对比学习得到的后继特征（经变量替换后）可构成满足三角不等式的时间距离，并用于强化学习中的规划与控制。
Temporal distances lie at the heart of many algorithms for planning, control, and reinforcement learning that involve reaching goals, allowing one to estimate the transit time between two states. However, prior attempts to define such temporal distances in stochastic settings have been stymied by an important limitation: these prior approaches do not satisfy the triangle inequality. This is not merely a definitional concern, but translates to an inability to generalize and find shortest paths. In this paper, we build on prior work in contrastive learning and quasimetrics to show how successor features learned by contrastive learning (after a change of variables) form a temporal distance that does satisfy the triangle inequality, even in stochastic settings. Importantly, this temporal distance is computationally efficient to estimate, even in high-dimensional and stochastic settings. Experiments in controlled settings and benchmark suites demonstrate that an RL algorithm based on these new temporal distances exhibits combinatorial generalization (i.e., "stitching") and can sometimes learn more quickly than prior methods, including those based on quasimetrics.
Offline Reinforcement Learning with Implicit Q-Learning
Ilya Kostrikov, Ashvin Nair, Sergey Levine
AI 速览：提出隐式Q学习（IQL），通过状态条件上期望分位数估计最优动作价值，无需查询数据集外动作即可改进策略。
Offline reinforcement learning requires reconciling two conflicting aims: learning a policy that improves over the behavior policy that collected the dataset, while at the same time minimizing the deviation from the behavior policy so as to avoid errors due to distributional shift. This trade-off is critical, because most current offline reinforcement learning methods need to query the value of unseen actions during training to improve the policy, and therefore need to either constrain these actions to be in-distribution, or else regularize their values. We propose an offline RL method that never needs to evaluate actions outside of the dataset, but still enables the learned policy to improve substantially over the best behavior in the data through generalization. The main insight in our work is that, instead of evaluating unseen actions from the latest policy, we can approximate the policy improvement step implicitly by treating the state value function as a random variable, with randomness determined by the action (while still integrating over the dynamics to avoid excessive optimism), and then taking a state conditional upper expectile of this random variable to estimate the value of the best actions in that state. This leverages the generalization capacity of the function approximator to estimate the value of the best available action at a given state without ever directly querying a Q-function with this unseen action. Our algorithm alternates between fitting this upper expectile value function and backing it up into a Q-function. Then, we extract the policy via advantage-weighted behavioral cloning. We dub our method implicit Q-learning (IQL). IQL demonstrates the state-of-the-art performance on D4RL, a standard benchmark for offline reinforcement learning. We also demonstrate that IQL achieves strong performance fine-tuning using online interaction after offline initialization.
Adversarial Policies: Attacking Deep Reinforcement Learning
Adam Gleave, Michael Dennis, Cody Wild, Neel Kant, Sergey Levine, Stuart Russell
AI 速览：作者展示在零和博弈中，仅通过选择对抗性策略制造自然观测，就能稳定击败经自博弈训练、对对手鲁棒的深度强化学习智能体。
Deep reinforcement learning (RL) policies are known to be vulnerable to adversarial perturbations to their observations, similar to adversarial examples for classifiers. However, an attacker is not usually able to directly modify another agent's observations. This might lead one to wonder: is it possible to attack an RL agent simply by choosing an adversarial policy acting in a multi-agent environment so as to create natural observations that are adversarial? We demonstrate the existence of adversarial policies in zero-sum games between simulated humanoid robots with proprioceptive observations, against state-of-the-art victims trained via self-play to be robust to opponents. The adversarial policies reliably win against the victims but generate seemingly random and uncoordinated behavior. We find that these policies are more successful in high-dimensional environments, and induce substantially different activations in the victim policy network than when the victim plays against a normal opponent. Videos are available at https://adversarialpolicies.github.io/.