Offline reinforcement learning (RL) defines the task of learning from a static logged dataset without continually interacting with the environment. The distribution shift between the learned policy and the behavior policy makes it necessary for the value function to stay conservative such that out-of-distribution (OOD) actions will not be severely overestimated. However, existing approaches, penalizing the unseen actions or regularizing with the behavior policy, are too pessimistic, which suppresses the generalization of the value function and hinders the performance improvement. This paper explores mild but enough conservatism for offline learning while not harming generalization. We propose Mildly Conservative Q-learning (MCQ), where OOD actions are actively trained by assigning them proper pseudo Q values. We theoretically show that MCQ induces a policy that behaves at least as well as the behavior policy and no erroneous overestimation will occur for OOD actions. Experimental results on the D4RL benchmarks demonstrate that MCQ achieves remarkable performance compared with prior work. Furthermore, MCQ shows superior generalization ability when transferring from offline to online, and significantly outperforms baselines. Our code is publicly available at https://github.com/dmksjfl/MCQ.

Effectively leveraging large, previously collected datasets in reinforcement learning (RL) is a key challenge for large-scale real-world applications. Offline RL algorithms promise to learn effective policies from previously-collected, static datasets without further interaction. However, in practice, offline RL presents a major challenge, and standard off-policy RL methods can fail due to overestimation of values induced by the distributional shift between the dataset and the learned policy, especially when training on complex and multi-modal data distributions. In this paper, we propose conservative Q-learning (CQL), which aims to address these limitations by learning a conservative Q-function such that the expected value of a policy under this Q-function lower-bounds its true value. We theoretically show that CQL produces a lower bound on the value of the current policy and that it can be incorporated into a policy learning procedure with theoretical improvement guarantees. In practice, CQL augments the standard Bellman error objective with a simple Q-value regularizer which is straightforward to implement on top of existing deep Q-learning and actor-critic implementations. On both discrete and continuous control domains, we show that CQL substantially outperforms existing offline RL methods, often learning policies that attain 2-5 times higher final return, especially when learning from complex and multi-modal data distributions.

Offline inverse reinforcement learning (IRL) aims to recover a reward function that explains expert behavior using only fixed demonstration data, without any additional online interaction. We propose BiCQL-ML, a policy-free offline IRL algorithm that jointly optimizes a reward function and a conservative Q-function in a bi-level framework, thereby avoiding explicit policy learning. The method alternates between (i) learning a conservative Q-function via Conservative Q-Learning (CQL) under the current reward, and (ii) updating the reward parameters to maximize the expected Q-values of expert actions while suppressing over-generalization to out-of-distribution actions. This procedure can be viewed as maximum likelihood estimation under a soft value matching principle. We provide theoretical guarantees that BiCQL-ML converges to a reward function under which the expert policy is soft-optimal. Empirically, we show on standard offline RL benchmarks that BiCQL-ML improves both reward recovery and downstream policy performance compared to existing offline IRL baselines.

The field of Offline Reinforcement Learning (RL) aims to derive effective policies from pre-collected datasets without active environment interaction. While traditional offline RL algorithms like Conservative Q-Learning (CQL) and Implicit Q-Learning (IQL) have shown promise, they often face challenges in balancing exploration and exploitation, especially in environments with varying reward densities. The recently proposed Decision Transformer (DT) approach, which reframes offline RL as a sequence modelling problem, has demonstrated impressive results across various benchmarks. This paper presents a comparative study evaluating the performance of DT against traditional offline RL algorithms in dense and sparse reward settings for the ANT con-tinous control environment. Our research investigates how these algorithms perform when faced with different reward structures, examining their ability to learn effective policies and generalize across varying levels of feedback. Through empirical analysis in the ANT environment, we found that DTs showed less sensitivity to varying reward density compared to other methods and particularly excelled with medium-expert datasets in sparse reward scenarios. In contrast, traditional value-based methods like IQL showed improved performance in dense reward settings with high-quality data, while CQL offered balanced performance across different data qualities. Additionally, DTs exhibited lower variance in performance but required significantly more computational resources compared to traditional approaches. These findings suggest that sequence modelling approaches may be more suitable for scenarios with uncertain reward structures or mixed-quality data, while value-based methods remain competitive in settings with dense rewards and high-quality demonstrations.

Soft snake robots offer remarkable flexibility and adaptability in complex environments, yet their control remains challenging due to highly nonlinear dynamics. Existing model-based and bio-inspired controllers rely on simplified assumptions that limit performance. Deep reinforcement learning (DRL) has recently emerged as a promising alternative, but online training is often impractical because of costly and potentially damaging real-world interactions. Offline RL provides a safer option by leveraging pre-collected datasets, but it suffers from distribution shift, which degrades generalization to unseen scenarios. To overcome this challenge, we propose DiSA-IQL (Distribution-Shift-Aware Implicit Q-Learning), an extension of IQL that incorporates robustness modulation by penalizing unreliable state-action pairs to mitigate distribution shift. We evaluate DiSA-IQL on goal-reaching tasks across two settings: in-distribution and out-of-distribution evaluation. Simulation results show that DiSA-IQL consistently outperforms baseline models, including Behavior Cloning (BC), Conservative Q-Learning (CQL), and vanilla IQL, achieving higher success rates, smoother trajectories, and improved robustness. The codes are open-sourced to support reproducibility and to facilitate further research in offline RL for soft robot control.

The theoretical claims of producing lower bounds on Q values are not sufficient since there is no proof that the conservative Q values are anywhere near the true Q values. Just estimating Q 0 for positive rewards could give the same result at theorems 3.1, 3.2, and 3.3. Clearly, the algorithm is doing something smarter than this, but the current analysis does not characterize what the algorithm is doing. The gap expanding result is likely the strongest of the four theorems, but without doing the work to connect this back to why this will actually help performance it is still difficult to judge. Moreover, no comparison is made to the overestimation that would happen without the proposed algorithmic change.

All the reviewers were positively impressed by the rebuttal provided by the authors, which clarified many of their concerns. Their updated scores and the final decision to propose acceptance for the paper are based on the requirement that the authors will significantly change the paper integrating the insights presented in the rebuttal as well as clarifications of the few remaining points that have not be addressed yet (see R3).

Offline reinforcement learning (RL) defines the task of learning from a static logged dataset without continually interacting with the environment. The distribution shift between the learned policy and the behavior policy makes it necessary for the value function to stay conservative such that out-of-distribution (OOD) actions will not be severely overestimated. However, existing approaches, penalizing the unseen actions or regularizing with the behavior policy, are too pessimistic, which suppresses the generalization of the value function and hinders the performance improvement. This paper explores mild but enough conservatism for offline learning while not harming generalization. We propose Mildly Conservative Q-learning (MCQ), where OOD actions are actively trained by assigning them proper pseudo Q values.

Effectively leveraging large, previously collected datasets in reinforcement learn- ing (RL) is a key challenge for large-scale real-world applications. Offline RL algorithms promise to learn effective policies from previously-collected, static datasets without further interaction. However, in practice, offline RL presents a major challenge, and standard off-policy RL methods can fail due to overestimation of values induced by the distributional shift between the dataset and the learned policy, especially when training on complex and multi-modal data distributions. In this paper, we propose conservative Q-learning (CQL), which aims to address these limitations by learning a conservative Q-function such that the expected value of a policy under this Q-function lower-bounds its true value. We theoretically show that CQL produces a lower bound on the value of the current policy and that it can be incorporated into a policy learning procedure with theoretical improvement guarantees. In practice, CQL augments the standard Bellman error objective with a simple Q-value regularizer which is straightforward to implement on top of existing deep Q-learning and actor-critic implementations.

Mischief can happen when AI is let loose in the world, just like any technology. The examples of AI gone wrong are numerous, the most vivid in recent memory being the disastrously bad performance of Amazon's facial recognition technology, Rekognition, which had a propensity to erroneously match members of some ethnic groups with criminal mugshots to a disproportionate extent. Given the risk, how can society know if a technology has been adequately refined to a level where it is safe to deploy? "This is a really good question, and one we are actively working on," Sergey Levine, assistant professor with the University of California at Berkeley's department of electrical engineering and computer science, told ZDNet by email this week. Levine and colleagues have been working on an approach to machine learning where the decisions of a software program are subjected to a critique by another algorithm within the same program that acts adversarially.