Songyang Han

Sihong He, Songyang Han, Sanbao Su et al.

In real-world multi-agent reinforcement learning (MARL) applications, agents may not have perfect state information (e.g., due to inaccurate measurement or malicious attacks), which challenges the robustness of agents' policies. Though robustness is getting important in MARL deployment, little prior work has studied state uncertainties in MARL, neither in problem formulation nor algorithm design. Motivated by this robustness issue and the lack of corresponding studies, we study the problem of MARL with state uncertainty in this work. We provide the first attempt to the theoretical and empirical analysis of this challenging problem. We first model the problem as a Markov Game with state perturbation adversaries (MG-SPA) by introducing a set of state perturbation adversaries into a Markov Game. We then introduce robust equilibrium (RE) as the solution concept of an MG-SPA. We conduct a fundamental analysis regarding MG-SPA such as giving conditions under which such a robust equilibrium exists. Then we propose a robust multi-agent Q-learning (RMAQ) algorithm to find such an equilibrium, with convergence guarantees. To handle high-dimensional state-action space, we design a robust multi-agent actor-critic (RMAAC) algorithm based on an analytical expression of the policy gradient derived in the paper. Our experiments show that the proposed RMAQ algorithm converges to the optimal value function; our RMAAC algorithm outperforms several MARL and robust MARL methods in multiple multi-agent environments when state uncertainty is present. The source code is public on \url{https://github.com/sihongho/robust_marl_with_state_uncertainty}.

Hao Zhang, Mingjie Liu, Shaokun Zhang et al.

Multi-turn LLM agents are increasingly important for solving complex, interactive tasks, and reinforcement learning (RL) is a key ingredient for improving their long-horizon behavior. However, RL training requires generating large numbers of sandboxed rollout trajectories, and existing infrastructures often couple rollout orchestration with the training loop, making systems hard to migrate and maintain. Under the rollout-as-a-service philosophy, we present ProRL Agent , a scalable infrastructure that serves the full agentic rollout lifecycle through an API service. ProRL Agent also provides standardized and extensible sandbox environments that support diverse agentic tasks in rootless HPC settings. We validate ProRL Agent through RL training on software engineering, math, STEM, and coding tasks. ProRL Agent is open-sourced and integrated as part of NVIDIA NeMo Gym.

Sanbao Su, Yiming Li, Sihong He et al.

Sharing information between connected and autonomous vehicles (CAVs) fundamentally improves the performance of collaborative object detection for self-driving. However, CAVs still have uncertainties on object detection due to practical challenges, which will affect the later modules in self-driving such as planning and control. Hence, uncertainty quantification is crucial for safety-critical systems such as CAVs. Our work is the first to estimate the uncertainty of collaborative object detection. We propose a novel uncertainty quantification method, called Double-M Quantification, which tailors a moving block bootstrap (MBB) algorithm with direct modeling of the multivariant Gaussian distribution of each corner of the bounding box. Our method captures both the epistemic uncertainty and aleatoric uncertainty with one inference pass based on the offline Double-M training process. And it can be used with different collaborative object detectors. Through experiments on the comprehensive collaborative perception dataset, we show that our Double-M method achieves more than 4X improvement on uncertainty score and more than 3% accuracy improvement, compared with the state-of-the-art uncertainty quantification methods. Our code is public on https://coperception.github.io/double-m-quantification.

Sanbao Su, Songyang Han, Yiming Li et al.

Object detection and multiple object tracking (MOT) are essential components of self-driving systems. Accurate detection and uncertainty quantification are both critical for onboard modules, such as perception, prediction, and planning, to improve the safety and robustness of autonomous vehicles. Collaborative object detection (COD) has been proposed to improve detection accuracy and reduce uncertainty by leveraging the viewpoints of multiple agents. However, little attention has been paid to how to leverage the uncertainty quantification from COD to enhance MOT performance. In this paper, as the first attempt to address this challenge, we design an uncertainty propagation framework called MOT-CUP. Our framework first quantifies the uncertainty of COD through direct modeling and conformal prediction, and propagates this uncertainty information into the motion prediction and association steps. MOT-CUP is designed to work with different collaborative object detectors and baseline MOT algorithms. We evaluate MOT-CUP on V2X-Sim, a comprehensive collaborative perception dataset, and demonstrate a 2% improvement in accuracy and a 2.67X reduction in uncertainty compared to the baselines, e.g. SORT and ByteTrack. In scenarios characterized by high occlusion levels, our MOT-CUP demonstrates a noteworthy $4.01\%$ improvement in accuracy. MOT-CUP demonstrates the importance of uncertainty quantification in both COD and MOT, and provides the first attempt to improve the accuracy and reduce the uncertainty in MOT based on COD through uncertainty propagation. Our code is public on https://coperception.github.io/MOT-CUP/.

Songyang Han, Sanbao Su, Sihong He et al.

Various methods for Multi-Agent Reinforcement Learning (MARL) have been developed with the assumption that agents' policies are based on accurate state information. However, policies learned through Deep Reinforcement Learning (DRL) are susceptible to adversarial state perturbation attacks. In this work, we propose a State-Adversarial Markov Game (SAMG) and make the first attempt to investigate different solution concepts of MARL under state uncertainties. Our analysis shows that the commonly used solution concepts of optimal agent policy and robust Nash equilibrium do not always exist in SAMGs. To circumvent this difficulty, we consider a new solution concept called robust agent policy, where agents aim to maximize the worst-case expected state value. We prove the existence of robust agent policy for finite state and finite action SAMGs. Additionally, we propose a Robust Multi-Agent Adversarial Actor-Critic (RMA3C) algorithm to learn robust policies for MARL agents under state uncertainties. Our experiments demonstrate that our algorithm outperforms existing methods when faced with state perturbations and greatly improves the robustness of MARL policies. Our code is public on https://songyanghan.github.io/what_is_solution/.

Jiangwei Wang, Shuo Yang, Ziyan An et al.

Reward design is a key component of deep reinforcement learning, yet some tasks and designer's objectives may be unnatural to define as a scalar cost function. Among the various techniques, formal methods integrated with DRL have garnered considerable attention due to their expressiveness and flexibility to define the reward and requirements for different states and actions of the agent. However, how to leverage Signal Temporal Logic (STL) to guide multi-agent reinforcement learning reward design remains unexplored. Complex interactions, heterogeneous goals and critical safety requirements in multi-agent systems make this problem even more challenging. In this paper, we propose a novel STL-guided multi-agent reinforcement learning framework. The STL requirements are designed to include both task specifications according to the objective of each agent and safety specifications, and the robustness values of the STL specifications are leveraged to generate rewards. We validate the advantages of our method through empirical studies. The experimental results demonstrate significant reward performance improvements compared to MARL without STL guidance, along with a remarkable increase in the overall safety rate of the multi-agent systems.

Manan Suri, Xiangci Li, Mehdi Shojaie et al. · amazon-science

Current AI-powered code assistance tools often struggle with poorly-defined problem statements that lack sufficient task context and requirements specification. Recent analysis of software engineering agents reveals that failures on such underspecified requests are highly correlated with longer trajectories involving either over-exploration or repeated attempts at applying the same fix without proper evolution or testing, leading to suboptimal outcomes across software development tasks. We introduce CodeScout, a contextual query refinement approach that systematically converts underspecified user requests into comprehensive, actionable problem statements through lightweight pre-exploration of the target codebase. Our key innovation is demonstrating that structured analysis before task execution can supplement existing agentic capabilities without requiring any modifications to their underlying scaffolds. CodeScout performs targeted context scoping, conducts multi-perspective analysis examining potential fixes and exploration opportunities, then synthesizes these insights into enhanced problem statements with reproduction steps, expected behaviors, and targeted exploration hints. This pre-exploration directly addresses the identified failure patterns by reducing non-converging agent trajectories while clarifying user intent in natural language space. We evaluate CodeScout using state-of-the-art agentic scaffolds and language models on SWEBench-Verified, demonstrating a 20\% improvement in resolution rates with up to 27 additional issues resolved compared to the default baseline method. Our results suggest that systematic query refinement through contextual analysis represents a promising direction for enhancing AI code assistance capabilities.

Songyang Han, Shanglin Zhou, Lynn Pepin et al.

The recent advancements in wireless technology enable connected autonomous vehicles (CAVs) to gather data via vehicle-to-vehicle (V2V) communication, such as processed LIDAR and camera data from other vehicles. In this work, we design an integrated information sharing and safe multi-agent reinforcement learning (MARL) framework for CAVs, to take advantage of the extra information when making decisions to improve traffic efficiency and safety. We first use weight pruned convolutional neural networks (CNN) to process the raw image and point cloud LIDAR data locally at each autonomous vehicle, and share CNN-output data with neighboring CAVs. We then design a safe actor-critic algorithm that utilizes both a vehicle's local observation and the information received via V2V communication to explore an efficient behavior planning policy with safety guarantees. Using the CARLA simulator for experiments, we show that our approach improves the CAV system's efficiency in terms of average velocity and comfort under different CAV ratios and different traffic densities. We also show that our approach avoids the execution of unsafe actions and always maintains a safe distance from other vehicles. We construct an obstacle-at-corner scenario to show that the shared vision can help CAVs to observe obstacles earlier and take action to avoid traffic jams.

Zhili Zhang, Songyang Han, Jiangwei Wang et al.

Communication technologies enable coordination among connected and autonomous vehicles (CAVs). However, it remains unclear how to utilize shared information to improve the safety and efficiency of the CAV system in dynamic and complicated driving scenarios. In this work, we propose a framework of constrained multi-agent reinforcement learning (MARL) with a parallel Safety Shield for CAVs in challenging driving scenarios that includes unconnected hazard vehicles. The coordination mechanisms of the proposed MARL include information sharing and cooperative policy learning, with Graph Convolutional Network (GCN)-Transformer as a spatial-temporal encoder that enhances the agent's environment awareness. The Safety Shield module with Control Barrier Functions (CBF)-based safety checking protects the agents from taking unsafe actions. We design a constrained multi-agent advantage actor-critic (CMAA2C) algorithm to train safe and cooperative policies for CAVs. With the experiment deployed in the CARLA simulator, we verify the performance of the safety checking, spatial-temporal encoder, and coordination mechanisms designed in our method by comparative experiments in several challenging scenarios with unconnected hazard vehicles. Results show that our proposed methodology significantly increases system safety and efficiency in challenging scenarios.

Jiangwei Wang, Lili Su, Songyang Han et al.

Hybrid traffic which involves both autonomous and human-driven vehicles would be the norm of the autonomous vehicles practice for a while. On the one hand, unlike autonomous vehicles, human-driven vehicles could exhibit sudden abnormal behaviors such as unpredictably switching to dangerous driving modes, putting its neighboring vehicles under risks; such undesired mode switching could arise from numbers of human driver factors, including fatigue, drunkenness, distraction, aggressiveness, etc. On the other hand, modern vehicle-to-vehicle communication technologies enable the autonomous vehicles to efficiently and reliably share the scarce run-time information with each other. In this paper, we propose, to the best of our knowledge, the first efficient algorithm that can (1) significantly improve trajectory prediction by effectively fusing the run-time information shared by surrounding autonomous vehicles, and can (2) accurately and quickly detect abnormal human driving mode switches or abnormal driving behavior with formal assurance without hurting human drivers privacy. To validate our proposed algorithm, we first evaluate our proposed trajectory predictor on NGSIM and Argoverse datasets and show that our proposed predictor outperforms the baseline methods. Then through extensive experiments on SUMO simulator, we show that our proposed algorithm has great detection performance in both highway and urban traffic. The best performance achieves detection rate of 97.3%, average detection delay of 1.2s, and 0 false alarm.

Binfeng Xu, Hao Zhang, Shaokun Zhang et al.

Reinforcement learning for language agents increasingly depends on custom harnesses that manage long-running context, multi-turn tool use and multi-agent orchestration. However, porting these harnesses into RL environment interfaces remains difficult and often loses important training signals. We bridge this gap with polar, a rollout framework for scalable asynchronous RL over arbitrary agent harnesses. Polar treats the agent harness as a black box: it proxies LLM API calls, records token-level model interactions, and reconstructs token-faithful trajectories for training. Each rollout node efficiently manages runtime prewarming, agent execution, trajectory reconstruction, and evaluation in parallel, exposing asynchronous service endpoints that can be consumed by independent trainers at scale. This decoupled design makes Polar agnostic to agent harnesses, training infrastructure, and RL algorithms while improving compute utilization for long-running agent workloads. We validate polar by training agents on software-engineering tasks with popular coding harnesses. Using simple GRPO, polar improves Qwen3.5-4B by 22.6, 4.8, 0.6 and 6.2 points on SWE-Bench Verified with the Codex, Claude Code, Qwen Code and Pi harnesses, respectively. We further demonstrate Polar for offline data generation over custom harnesses and ablate trajectory reconstruction strategies. Polar rewrites its preceding work, Prorl Agent, and has been registered as one of NeMo Gym environments.

Kewei Zhang, Jin Wang, Sensen Gao et al.

End-to-end autonomous driving via Vision-Language-Action (VLA) models demands a precarious balance between high-fidelity trajectory planning and efficient inference. Existing paradigms typically fall short: autoregressive (AR) VLAs are memory-bandwidth-bound on edge hardware and prone to exposure-bias drift, while full-sequence diffusion models preclude KV-cache reuse and suffer from "logical leakage" that violates the fundamental perceive-then-plan causality. We present Fast-dDrive, a block-diffusion VLA that performs bidirectional refinement within semantic units while enforcing strict causal ordering across them. Leveraging the observation that driving VLAs often emit structured JSON-like outputs, Fast-dDrive freezes structural tokens into a section scaffold and employs a section-aware training recipe that prioritizes safety-critical planning. We further introduce Scaffold Speculative Decoding to achieve AR-equivalent quality at significantly higher throughput. Finally, we propose a low-overhead test-time scaling scheme: by forking $N$ stochastic trajectory rollouts from a single shared-prefix KV cache and averaging them, we effectively suppress prediction variance at a fractional computational cost. Empirical results demonstrate that Fast-dDrive redefines the speed-accuracy frontier for driving agents. On the WOD-E2E test set, Fast-dDrive achieves SOTA ADE@3s and ADE@5s, alongside the highest RFS among diffusion-based VLAs; on nuScenes, it reduces average L2 error to $0.32$m (a $22\%$ improvement). When integrated with SGLang, our framework delivers $12\times$ throughput speedup over the AR baseline, narrowing the gap between high-capacity VLAs and the efficiency demands of real-time on-vehicle deployment.

Yuan Zhuang, Yuexin Bian, Sihong He et al.

Scaling critic capacity is a promising direction for enhancing off-policy reinforcement learning (RL). However, larger critics are prone to overfitting and unstable in replay-buffer-based bootstrap training. This paper leverages Low-Rank Adaptation (LoRA) as a structural-sparsity regularizer for off-policy critics. Our approach freezes randomly initialized base matrices and solely optimizes low-rank adapters, thereby constraining critic updates to a low-dimensional subspace. Built on top of SimbaV2, we further develop a LoRA formulation, compatible with SimbaV2, that preserves its hyperspherical normalization geometry under frozen-backbone training. We evaluate our method with SAC and FastTD3 on DeepMind Control locomotion and IsaacLab robotics benchmarks. LoRA consistently achieves lower critic loss during training and stronger policy performance. Extensive experiments demonstrate that adaptive low-rank updates provide a simple, scalable, and effective structural regularization for critic learning in off-policy RL.

Yifeng Ding, Hung Le, Songyang Han et al.

Training Large Language Models (LLMs) for multi-turn Tool-Integrated Reasoning (TIR) - where models iteratively reason, generate code, and verify through execution - remains challenging for existing reinforcement learning (RL) approaches. Current RL methods, exemplified by Group Relative Policy Optimization (GRPO), suffer from coarse-grained, trajectory-level rewards that provide insufficient learning signals for complex multi-turn interactions, leading to training stagnation. To address this issue, we propose Group Turn Policy Optimization (GTPO), a novel RL algorithm specifically designed for training LLMs on multi-turn TIR tasks. GTPO introduces three key innovations: (1) turn-level reward assignment that provides fine-grained feedback for individual turns, (2) return-based advantage estimation where normalized discounted returns are calculated as advantages, and (3) self-supervised reward shaping that exploits self-supervision signals from generated code to densify sparse binary outcome-based rewards. Our comprehensive evaluation demonstrates that GTPO outperforms GRPO by 3.0% on average across diverse reasoning benchmarks, establishing its effectiveness for advancing complex mathematical reasoning in the real world.

Songyang Han, Shanglin Zhou, Jiangwei Wang et al.

The recent advancements in wireless technology enable connected autonomous vehicles (CAVs) to gather information about their environment by vehicle-to-vehicle (V2V) communication. In this work, we design an information-sharing-based multi-agent reinforcement learning (MARL) framework for CAVs, to take advantage of the extra information when making decisions to improve traffic efficiency and safety. The safe actor-critic algorithm we propose has two new techniques: the truncated Q-function and safe action mapping. The truncated Q-function utilizes the shared information from neighboring CAVs such that the joint state and action spaces of the Q-function do not grow in our algorithm for a large-scale CAV system. We prove the bound of the approximation error between the truncated-Q and global Q-functions. The safe action mapping provides a provable safety guarantee for both the training and execution based on control barrier functions. Using the CARLA simulator for experiments, we show that our approach can improve the CAV system's efficiency in terms of average velocity and comfort under different CAV ratios and different traffic densities. We also show that our approach avoids the execution of unsafe actions and always maintains a safe distance from other vehicles. We construct an obstacle-at-corner scenario to show that the shared vision can help CAVs to observe obstacles earlier and take action to avoid traffic jams.