In recent years, the exceptional performance of generative models in generative tasks has sparked significant interest in their integration into decision-making processes. Due to their ability to handle complex data distributions and their strong model capacity, generative models can be effectively incorporated into decision-making systems by generating trajectories that guide agents toward high-reward state-action regions or intermediate sub-goals. This paper presents a comprehensive review of the application of generative models in decision-making tasks. We classify seven fundamental types of generative models: energy-based models, generative adversarial networks, variational autoencoders, normalizing flows, diffusion models, generative flow networks, and autoregressive models. Regarding their applications, we categorize their functions into three main roles: controllers, modelers and optimizers, and discuss how each role contributes to decision-making. Furthermore, we examine the deployment of these models across five critical real-world decision-making scenarios. Finally, we summarize the strengths and limitations of current approaches and propose three key directions for advancing next-generation generative directive models: high-performance algorithms, large-scale generalized decision-making models, and self-evolving and adaptive models.

Reinforcement learning with verifiable outcome rewards (RLVR) has effectively scaled up chain-of-thought (CoT) reasoning in large language models (LLMs). Yet, its efficacy in training vision-language model (VLM) agents for goal-directed action reasoning in visual environments is less established. This work investigates this problem through extensive experiments on complex card games, such as 24 points, and embodied tasks from ALFWorld. We find that when rewards are based solely on action outcomes, RL fails to incentivize CoT reasoning in VLMs, instead leading to a phenomenon we termed thought collapse, characterized by a rapid loss of diversity in the agent's thoughts, state-irrelevant and incomplete reasoning, and subsequent invalid actions, resulting in negative rewards. To counteract thought collapse, we highlight the necessity of process guidance and propose an automated corrector that evaluates and refines the agent's reasoning at each RL step. This simple and scalable GTR (Guided Thought Reinforcement) framework trains reasoning and action simultaneously without the need for dense, per-step human labeling. Our experiments demonstrate that GTR significantly enhances the performance and generalization of the LLaVA-7b model across various visual environments, achieving 3-5 times higher task success rates compared to SoTA models with notably smaller model sizes.

Radioguided surgery, such as sentinel lymph node biopsy, relies on the precise localization of radioactive targets by non-imaging gamma/beta detectors. Manual radioactive target detection based on visual display or audible indication of gamma level is highly dependent on the ability of the surgeon to track and interpret the spatial information. This paper presents a learning-based method to realize the autonomous radiotracer detection in robot-assisted surgeries by navigating the probe to the radioactive target. We proposed novel hybrid approach that combines deep reinforcement learning (DRL) with adaptive robotic scanning. The adaptive grid-based scanning could provide initial direction estimation while the DRL-based agent could efficiently navigate to the target utilising historical data. Simulation experiments demonstrate a 95% success rate, and improved efficiency and robustness compared to conventional techniques. Real-world evaluation on the da Vinci Research Kit (dVRK) further confirms the feasibility of the approach, achieving an 80% success rate in radiotracer detection. This method has the potential to enhance consistency, reduce operator dependency, and improve procedural accuracy in radioguided surgeries.

On-robot Reinforcement Learning is a promising approach to train embodiment-aware policies for legged robots. However, the computational constraints of real-time learning on robots pose a significant challenge. We present a framework for efficiently learning quadruped locomotion in just 8 minutes of raw real-time training utilizing the sample efficiency and minimal computational overhead of the new off-policy algorithm CrossQ. We investigate two control architectures: Predicting joint target positions for agile, high-speed locomotion and Central Pattern Generators for stable, natural gaits. While prior work focused on learning simple forward gaits, our framework extends on-robot learning to omnidirectional locomotion. We demonstrate the robustness of our approach in different indoor and outdoor environments.

In recent years, research on humanoid robots has garnered significant attention, particularly in reinforcement learning based control algorithms, which have achieved major breakthroughs. Compared to traditional model-based control algorithms, reinforcement learning based algorithms demonstrate substantial advantages in handling complex tasks. Leveraging the large-scale parallel computing capabilities of GPUs, contemporary humanoid robots can undergo extensive parallel training in simulated environments. A physical simulation platform capable of large-scale parallel training is crucial for the development of humanoid robots. As one of the most complex robot forms, humanoid robots typically possess intricate mechanical structures, encompassing numerous series and parallel mechanisms. However, many reinforcement learning based humanoid robot control algorithms currently employ open-loop topologies during training, deferring the conversion to series-parallel structures until the sim2real phase. This approach is primarily due to the limitations of physics engines, as current GPU-based physics engines often only support open-loop topologies or have limited capabilities in simulating multi-rigid-body closed-loop topologies. For enabling reinforcement learning-based humanoid robot control algorithms to train in large-scale parallel environments, we propose a novel training method LiPS. By incorporating multi-rigid-body dynamics modeling in the simulation environment, we significantly reduce the sim2real gap and the difficulty of converting to parallel structures during model deployment, thereby robustly supporting large-scale reinforcement learning for humanoid robots.

There exist applications of reinforcement learning like medicine where policies need to be ''interpretable'' by humans. User studies have shown that some policy classes might be more interpretable than others. However, it is costly to conduct human studies of policy interpretability. Furthermore, there is no clear definition of policy interpretabiliy, i.e., no clear metrics for interpretability and thus claims depend on the chosen definition. We tackle the problem of empirically evaluating policies interpretability without humans. Despite this lack of clear definition, researchers agree on the notions of ''simulatability'': policy interpretability should relate to how humans understand policy actions given states. To advance research in interpretable reinforcement learning, we contribute a new methodology to evaluate policy interpretability. This new methodology relies on proxies for simulatability that we use to conduct a large-scale empirical evaluation of policy interpretability. We use imitation learning to compute baseline policies by distilling expert neural networks into small programs. We then show that using our methodology to evaluate the baselines interpretability leads to similar conclusions as user studies. We show that increasing interpretability does not necessarily reduce performances and can sometimes increase them. We also show that there is no policy class that better trades off interpretability and performance across tasks making it necessary for researcher to have methodologies for comparing policies interpretability.

Quadcopter attitude control involves two tasks: smooth attitude tracking and aggressive stabilization from arbitrary states. Although both can be formulated as tracking problems, their distinct state spaces and control strategies complicate a unified reward function. We propose a multitask deep reinforcement learning framework that leverages parallel simulation with IsaacGym and a Graph Convolutional Network (GCN) policy to address both tasks effectively. Our multitask Soft Actor-Critic (SAC) approach achieves faster, more reliable learning and higher sample efficiency than single-task methods. We validate its real-world applicability by deploying the learned policy - a compact two-layer network with 24 neurons per layer - on a Pixhawk flight controller, achieving 400 Hz control without extra computational resources. We provide our code at https://github.com/robot-perception-group/GraphMTSAC\_UAV/.

Advancing safe autonomous systems through reinforcement learning (RL) requires robust benchmarks to evaluate performance, analyze methods, and assess agent competencies. Humans primarily rely on embodied visual perception to safely navigate and interact with their surroundings, making it a valuable capability for RL agents. However, existing vision-based 3D benchmarks only consider simple navigation tasks. To address this shortcoming, we introduce \textbf{HASARD}, a suite of diverse and complex tasks to $\textbf{HA}$rness $\textbf{SA}$fe $\textbf{R}$L with $\textbf{D}$oom, requiring strategic decision-making, comprehending spatial relationships, and predicting the short-term future. HASARD features three difficulty levels and two action spaces. An empirical evaluation of popular baseline methods demonstrates the benchmark's complexity, unique challenges, and reward-cost trade-offs. Visualizing agent navigation during training with top-down heatmaps provides insight into a method's learning process. Incrementally training across difficulty levels offers an implicit learning curriculum. HASARD is the first safe RL benchmark to exclusively target egocentric vision-based learning, offering a cost-effective and insightful way to explore the potential and boundaries of current and future safe RL methods. The environments and baseline implementations are open-sourced at https://sites.google.com/view/hasard-bench/.

Pok\'emon Red, a classic Game Boy JRPG, presents significant challenges as a testbed for agents, including multi-tasking, long horizons of tens of thousands of steps, hard exploration, and a vast array of potential policies. We introduce a simplistic environment and a Deep Reinforcement Learning (DRL) training methodology, demonstrating a baseline agent that completes an initial segment of the game up to completing Cerulean City. Our experiments include various ablations that reveal vulnerabilities in reward shaping, where agents exploit specific reward signals. We also discuss limitations and argue that games like Pok\'emon hold strong potential for future research on Large Language Model agents, hierarchical training algorithms, and advanced exploration methods. Source Code: https://github.com/MarcoMeter/neroRL/tree/poke_red

Reinforcement learning (RL) has demonstrated the ability to maintain the plasticity of the policy throughout short-term training in aerial robot control. However, these policies have been shown to loss of plasticity when extended to long-term learning in non-stationary environments. For example, the standard proximal policy optimization (PPO) policy is observed to collapse in long-term training settings and lead to significant control performance degradation. To address this problem, this work proposes a cost-aware framework that uses a retrospective cost mechanism (RECOM) to balance rewards and losses in RL training with a non-stationary environment. Using a cost gradient relation between rewards and losses, our framework dynamically updates the learning rate to actively train the control policy in a disturbed wind environment. Our experimental results show that our framework learned a policy for the hovering task without policy collapse in variable wind conditions and has a successful result of 11.29% less dormant units than L2 regularization with PPO.