Decision-making in complex, continuous multi-task environments is often hindered by the difficulty of obtaining accurate models for planning and the inefficiency of learning purely from trial and error. While precise environment dynamics may be hard to specify, human experts can often provide high-fidelity abstractions that capture the essential high-level structure of a task and user preferences in the target environment. Existing hierarchical approaches often target discrete settings and do not generalize across tasks. We propose a hierarchical reinforcement learning approach that addresses these limitations by dynamically planning over the expert-specified abstraction to generate subgoals to learn a goal-conditioned policy. To overcome the challenges of learning under sparse rewards, we shape the reward based on the optimal state value in the abstract model. This structured decision-making process enhances sample efficiency and facilitates zero-shot generalization. Our empirical evaluation on a suite of procedurally generated continuous control environments demonstrates that our approach outperforms existing hierarchical reinforcement learning methods in terms of sample efficiency, task completion rate, scalability to complex tasks, and generalization to novel scenarios.
Designing reward functions for continuous-control robotics often leads to subtle misalignments or reward hacking, especially in complex tasks. Preference-based RL mitigates some of these pitfalls by learning rewards from comparative feedback rather than hand-crafted signals, yet scaling human annotations remains challenging. Recent work uses Vision-Language Models (VLMs) to automate preference labeling, but a single final-state image generally fails to capture the agent's full motion. In this paper, we present a two-part solution that both improves feedback accuracy and better aligns reward learning with the agent's policy. First, we overlay trajectory sketches on final observations to reveal the path taken, allowing VLMs to provide more reliable preferences-improving preference accuracy by approximately 15-20% in metaworld tasks. Second, we regularize reward learning by incorporating the agent's performance, ensuring that the reward model is optimized based on data generated by the current policy; this addition boosts episode returns by 20-30% in locomotion tasks. Empirical studies on metaworld demonstrate that our method achieves, for instance, around 70-80% success rate in all tasks, compared to below 50% for standard approaches. These results underscore the efficacy of combining richer visual representations with agent-aware reward regularization.
Style transfer involves transferring the style from a reference image to the content of a target image. Recent advancements in LoRA-based (Low-Rank Adaptation) methods have shown promise in effectively capturing the style of a single image. However, these approaches still face significant challenges such as content inconsistency, style misalignment, and content leakage. In this paper, we comprehensively analyze the limitations of the standard diffusion parameterization, which learns to predict noise, in the context of style transfer. To address these issues, we introduce ConsisLoRA, a LoRA-based method that enhances both content and style consistency by optimizing the LoRA weights to predict the original image rather than noise. We also propose a two-step training strategy that decouples the learning of content and style from the reference image. To effectively capture both the global structure and local details of the content image, we introduce a stepwise loss transition strategy. Additionally, we present an inference guidance method that enables continuous control over content and style strengths during inference. Through both qualitative and quantitative evaluations, our method demonstrates significant improvements in content and style consistency while effectively reducing content leakage.
Dexterous telemanipulation critically relies on the continuous and stable tracking of the human operator's commands to ensure robust operation. Vison-based tracking methods are widely used but have low stability due to anomalies such as occlusions, inadequate lighting, and loss of sight. Traditional filtering, regression, and interpolation methods are commonly used to compensate for explicit information such as angles and positions. These approaches are restricted to low-dimensional data and often result in information loss compared to the original high-dimensional image and video data. Recent advances in diffusion-based approaches, which can operate on high-dimensional data, have achieved remarkable success in video reconstruction and generation. However, these methods have not been fully explored in continuous control tasks in robotics. This work introduces the Diffusion-Enhanced Telemanipulation (DET) framework, which incorporates the Frame-Difference Detection (FDD) technique to identify and segment anomalies in video streams. These anomalous clips are replaced after reconstruction using diffusion models, ensuring robust telemanipulation performance under challenging visual conditions. We validated this approach in various anomaly scenarios and compared it with the baseline methods. Experiments show that DET achieves an average RMSE reduction of 17.2% compared to the cubic spline and 51.1% compared to FFT-based interpolation for different occlusion durations.
In this paper, we study offline preference-based reinforcement learning (PbRL), where learning is based on pre-collected preference feedback over pairs of trajectories. While offline PbRL has demonstrated remarkable empirical success, existing theoretical approaches face challenges in ensuring conservatism under uncertainty, requiring computationally intractable confidence set constructions. We address this limitation by proposing Adversarial Preference-based Policy Optimization (APPO), a computationally efficient algorithm for offline PbRL that guarantees sample complexity bounds without relying on explicit confidence sets. By framing PbRL as a two-player game between a policy and a model, our approach enforces conservatism in a tractable manner. Using standard assumptions on function approximation and bounded trajectory concentrability, we derive a sample complexity bound. To our knowledge, APPO is the first offline PbRL algorithm to offer both statistical efficiency and practical applicability. Experimental results on continuous control tasks demonstrate that APPO effectively learns from complex datasets, showing comparable performance with existing state-of-the-art methods.
Practitioners designing reinforcement learning policies face a fundamental challenge: translating intended behavioral objectives into representative reward functions. This challenge stems from behavioral intent requiring simultaneous achievement of multiple competing objectives, typically addressed through labor-intensive linear reward composition that yields brittle results. Consider the ubiquitous robotics scenario where performance maximization directly conflicts with energy conservation. Such competitive dynamics are resistant to simple linear reward combinations. In this paper, we present the concept of objective fulfillment upon which we build Fulfillment Priority Logic (FPL). FPL allows practitioners to define logical formula representing their intentions and priorities within multi-objective reinforcement learning. Our novel Balanced Policy Gradient algorithm leverages FPL specifications to achieve up to 500\% better sample efficiency compared to Soft Actor Critic. Notably, this work constitutes the first implementation of non-linear utility scalarization design, specifically for continuous control problems.
On-policy reinforcement learning algorithms use the most recently learned policy to interact with the environment and update it using the latest gathered trajectories, making them well-suited for adapting to non-stationary environments where dynamics change over time. However, previous studies show that they struggle to maintain plasticity$\unicode{x2013}$the ability of neural networks to adjust their synaptic connections$\unicode{x2013}$with overfitting identified as the primary cause. To address this, we present the first application of evidential learning in an on-policy reinforcement learning setting: $\textit{Evidential Proximal Policy Optimization (EPPO)}$. EPPO incorporates all sources of error in the critic network's approximation$\unicode{x2013}$i.e., the baseline function in advantage calculation$\unicode{x2013}$by modeling the epistemic and aleatoric uncertainty contributions to the approximation's total variance. We achieve this by using an evidential neural network, which serves as a regularizer to prevent overfitting. The resulting probabilistic interpretation of the advantage function enables optimistic exploration, thus maintaining the plasticity. Through experiments on non-stationary continuous control tasks, where the environment dynamics change at regular intervals, we demonstrate that EPPO outperforms state-of-the-art on-policy reinforcement learning variants in both task-specific and overall return.
In reinforcement learning (RL), world models serve as internal simulators, enabling agents to predict environment dynamics and future outcomes in order to make informed decisions. While previous approaches leveraging discrete latent spaces, such as DreamerV3, have demonstrated strong performance in discrete action settings and visual control tasks, their comparative performance in state-based continuous control remains underexplored. In contrast, methods with continuous latent spaces, such as TD-MPC2, have shown notable success in state-based continuous control benchmarks. In this paper, we demonstrate that modeling discrete latent states has benefits over continuous latent states and that discrete codebook encodings are more effective representations for continuous control, compared to alternative encodings, such as one-hot and label-based encodings. Based on these insights, we introduce DCWM: Discrete Codebook World Model, a self-supervised world model with a discrete and stochastic latent space, where latent states are codes from a codebook. We combine DCWM with decision-time planning to get our model-based RL algorithm, named DC-MPC: Discrete Codebook Model Predictive Control, which performs competitively against recent state-of-the-art algorithms, including TD-MPC2 and DreamerV3, on continuous control benchmarks. See our project website www.aidanscannell.com/dcmpc.
Sequential decision-making in high-dimensional continuous action spaces, particularly in stochastic environments, faces significant computational challenges. We explore this challenge in the traditional offline RL setting, where an agent must learn how to make decisions based on data collected through a stochastic behavior policy. We present Latent Macro Action Planner (L-MAP), which addresses this challenge by learning a set of temporally extended macro-actions through a state-conditional Vector Quantized Variational Autoencoder (VQ-VAE), effectively reducing action dimensionality. L-MAP employs a (separate) learned prior model that acts as a latent transition model and allows efficient sampling of plausible actions. During planning, our approach accounts for stochasticity in both the environment and the behavior policy by using Monte Carlo tree search (MCTS). In offline RL settings, including stochastic continuous control tasks, L-MAP efficiently searches over discrete latent actions to yield high expected returns. Empirical results demonstrate that L-MAP maintains low decision latency despite increased action dimensionality. Notably, across tasks ranging from continuous control with inherently stochastic dynamics to high-dimensional robotic hand manipulation, L-MAP significantly outperforms existing model-based methods and performs on-par with strong model-free actor-critic baselines, highlighting the effectiveness of the proposed approach in planning in complex and stochastic environments with high-dimensional action spaces.
Current Reinforcement Learning (RL) methods often suffer from sample-inefficiency, resulting from blind exploration strategies that neglect causal relationships among states, actions, and rewards. Although recent causal approaches aim to address this problem, they lack grounded modeling of reward-guided causal understanding of states and actions for goal-orientation, thus impairing learning efficiency. To tackle this issue, we propose a novel method named Causal Information Prioritization (CIP) that improves sample efficiency by leveraging factored MDPs to infer causal relationships between different dimensions of states and actions with respect to rewards, enabling the prioritization of causal information. Specifically, CIP identifies and leverages causal relationships between states and rewards to execute counterfactual data augmentation to prioritize high-impact state features under the causal understanding of the environments. Moreover, CIP integrates a causality-aware empowerment learning objective, which significantly enhances the agent's execution of reward-guided actions for more efficient exploration in complex environments. To fully assess the effectiveness of CIP, we conduct extensive experiments across 39 tasks in 5 diverse continuous control environments, encompassing both locomotion and manipulation skills learning with pixel-based and sparse reward settings. Experimental results demonstrate that CIP consistently outperforms existing RL methods across a wide range of scenarios.
In this era of AI revolution, massive investments in large-scale data-driven AI systems demand high-performance computing, consuming tremendous energy and resources. This trend raises new challenges in optimizing sustainability without sacrificing scalability or performance. Among the energy-efficient alternatives of the traditional Von Neumann architecture, neuromorphic computing and its Spiking Neural Networks (SNNs) are a promising choice due to their inherent energy efficiency. However, in some real-world application scenarios such as complex continuous control tasks, SNNs often lack the performance optimizations that traditional artificial neural networks have. Researchers have addressed this by combining SNNs with Deep Reinforcement Learning (DeepRL), yet scalability remains unexplored. In this paper, we extend our previous work on SpikeRL, which is a scalable and energy efficient framework for DeepRL-based SNNs for continuous control. In our initial implementation of SpikeRL framework, we depended on the population encoding from the Population-coded Spiking Actor Network (PopSAN) method for our SNN model and implemented distributed training with Message Passing Interface (MPI) through mpi4py. Also, further optimizing our model training by using mixed-precision for parameter updates. In our new SpikeRL framework, we have implemented our own DeepRL-SNN component with population encoding, and distributed training with PyTorch Distributed package with NCCL backend while still optimizing with mixed precision training. Our new SpikeRL implementation is 4.26X faster and 2.25X more energy efficient than state-of-the-art DeepRL-SNN methods. Our proposed SpikeRL framework demonstrates a truly scalable and sustainable solution for complex continuous control tasks in real-world applications.
Scaling up the model size and computation has brought consistent performance improvements in supervised learning. However, this lesson often fails to apply to reinforcement learning (RL) because training the model on non-stationary data easily leads to overfitting and unstable optimization. In response, we introduce SimbaV2, a novel RL architecture designed to stabilize optimization by (i) constraining the growth of weight and feature norm by hyperspherical normalization; and (ii) using a distributional value estimation with reward scaling to maintain stable gradients under varying reward magnitudes. Using the soft actor-critic as a base algorithm, SimbaV2 scales up effectively with larger models and greater compute, achieving state-of-the-art performance on 57 continuous control tasks across 4 domains. The code is available at https://dojeon-ai.github.io/SimbaV2.
A major challenge for reinforcement learning is automatically generating curricula to reduce training time or improve performance in some target task. We introduce SEBNs (Skill-Environment Bayesian Networks) which model a probabilistic relationship between a set of skills, a set of goals that relate to the reward structure, and a set of environment features to predict policy performance on (possibly unseen) tasks. We develop an algorithm that uses the inferred estimates of agent success from SEBN to weigh the possible next tasks by expected improvement. We evaluate the benefit of the resulting curriculum on three environments: a discrete gridworld, continuous control, and simulated robotics. The results show that curricula constructed using SEBN frequently outperform other baselines.
Most control techniques for prosthetic grasping focus on dexterous fingers control, but overlook the wrist motion. This forces the user to perform compensatory movements with the elbow, shoulder and hip to adapt the wrist for grasping. We propose a computer vision-based system that leverages the collaboration between the user and an automatic system in a shared autonomy framework, to perform continuous control of the wrist degrees of freedom in a prosthetic arm, promoting a more natural approach-to-grasp motion. Our pipeline allows to seamlessly control the prosthetic wrist to follow the target object and finally orient it for grasping according to the user intent. We assess the effectiveness of each system component through quantitative analysis and finally deploy our method on the Hannes prosthetic arm. Code and videos: https://hsp-iit.github.io/hannes-wrist-control.
Proximal policy optimization (PPO) is one of the most popular state-of-the-art on-policy algorithms that has become a standard baseline in modern reinforcement learning with applications in numerous fields. Though it delivers stable performance with theoretical policy improvement guarantees, high variance, and high sample complexity still remain critical challenges in on-policy algorithms. To alleviate these issues, we propose Hybrid-Policy Proximal Policy Optimization (HP3O), which utilizes a trajectory replay buffer to make efficient use of trajectories generated by recent policies. Particularly, the buffer applies the "first in, first out" (FIFO) strategy so as to keep only the recent trajectories to attenuate the data distribution drift. A batch consisting of the trajectory with the best return and other randomly sampled ones from the buffer is used for updating the policy networks. The strategy helps the agent to improve its capability on top of the most recent best performance and in turn reduce variance empirically. We theoretically construct the policy improvement guarantees for the proposed algorithm. HP3O is validated and compared against several baseline algorithms using multiple continuous control environments. Our code is available here.
We introduce a scaling strategy for Explicit Policy-Conditioned Value Functions (EPVFs) that significantly improves performance on challenging continuous-control tasks. EPVFs learn a value function V({\theta}) that is explicitly conditioned on the policy parameters, enabling direct gradient-based updates to the parameters of any policy. However, EPVFs at scale struggle with unrestricted parameter growth and efficient exploration in the policy parameter space. To address these issues, we utilize massive parallelization with GPU-based simulators, big batch sizes, weight clipping and scaled peturbations. Our results show that EPVFs can be scaled to solve complex tasks, such as a custom Ant environment, and can compete with state-of-the-art Deep Reinforcement Learning (DRL) baselines like Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC). We further explore action-based policy parameter representations from previous work and specialized neural network architectures to efficiently handle weight-space features, which have not been used in the context of DRL before.
Reinforcement learning has achieved significant milestones, but sample efficiency remains a bottleneck for real-world applications. Recently, CrossQ has demonstrated state-of-the-art sample efficiency with a low update-to-data (UTD) ratio of 1. In this work, we explore CrossQ's scaling behavior with higher UTD ratios. We identify challenges in the training dynamics, which are emphasized by higher UTD ratios. To address these, we integrate weight normalization into the CrossQ framework, a solution that stabilizes training, has been shown to prevent potential loss of plasticity and keeps the effective learning rate constant. Our proposed approach reliably scales with increasing UTD ratios, achieving competitive performance across 25 challenging continuous control tasks on the DeepMind Control Suite and Myosuite benchmarks, notably the complex dog and humanoid environments. This work eliminates the need for drastic interventions, such as network resets, and offers a simple yet robust pathway for improving sample efficiency and scalability in model-free reinforcement learning.
Reinforcement Learning (RL) has achieved remarkable success in various continuous control tasks, such as robot manipulation and locomotion. Different to mainstream RL which makes decisions at individual steps, recent studies have incorporated action repetition into RL, achieving enhanced action persistence with improved sample efficiency and superior performance. However, existing methods treat all action dimensions as a whole during repetition, ignoring variations among them. This constraint leads to inflexibility in decisions, which reduces policy agility with inferior effectiveness. In this work, we propose a novel repetition framework called SDAR, which implements Spatially Decoupled Action Repetition through performing closed-loop act-or-repeat selection for each action dimension individually. SDAR achieves more flexible repetition strategies, leading to an improved balance between action persistence and diversity. Compared to existing repetition frameworks, SDAR is more sample efficient with higher policy performance and reduced action fluctuation. Experiments are conducted on various continuous control scenarios, demonstrating the effectiveness of spatially decoupled repetition design proposed in this work.
The primary focus of offline reinforcement learning (RL) is to manage the risk of hazardous exploitation of out-of-distribution actions. An effective approach to achieve this goal is through behavior regularization, which augments conventional RL objectives by incorporating constraints that enforce the policy to remain close to the behavior policy. Nevertheless, existing literature on behavior-regularized RL primarily focuses on explicit policy parameterizations, such as Gaussian policies. Consequently, it remains unclear how to extend this framework to more advanced policy parameterizations, such as diffusion models. In this paper, we introduce BDPO, a principled behavior-regularized RL framework tailored for diffusion-based policies, thereby combining the expressive power of diffusion policies and the robustness provided by regularization. The key ingredient of our method is to calculate the Kullback-Leibler (KL) regularization analytically as the accumulated discrepancies in reverse-time transition kernels along the diffusion trajectory. By integrating the regularization, we develop an efficient two-time-scale actor-critic RL algorithm that produces the optimal policy while respecting the behavior constraint. Comprehensive evaluations conducted on synthetic 2D tasks and continuous control tasks from the D4RL benchmark validate its effectiveness and superior performance.
Model-based reinforcement learning algorithms that combine model-based planning and learned value/policy prior have gained significant recognition for their high data efficiency and superior performance in continuous control. However, we discover that existing methods that rely on standard SAC-style policy iteration for value learning, directly using data generated by the planner, often result in \emph{persistent value overestimation}. Through theoretical analysis and experiments, we argue that this issue is deeply rooted in the structural policy mismatch between the data generation policy that is always bootstrapped by the planner and the learned policy prior. To mitigate such a mismatch in a minimalist way, we propose a policy regularization term reducing out-of-distribution (OOD) queries, thereby improving value learning. Our method involves minimum changes on top of existing frameworks and requires no additional computation. Extensive experiments demonstrate that the proposed approach improves performance over baselines such as TD-MPC2 by large margins, particularly in 61-DoF humanoid tasks. View qualitative results at https://darthutopian.github.io/tdmpc_square/.