This work introduces a growable and modular neural network architecture that naturally avoids catastrophic forgetting and interference in continual reinforcement learning. The structure of each module allows the selective combination of previous policies along with its internal policy, accelerating the learning process on the current task. Unlike previous growing neural network approaches, we show that the number of parameters of the proposed approach grows linearly with respect to the number of tasks, and does not sacrifice plasticity to scale. Experiments conducted in benchmark continuous control and visual problems reveal that the proposed approach achieves greater knowledge transfer and performance than alternative methods.

We present Ring-lite, a Mixture-of-Experts (MoE)-based large language model optimized via reinforcement learning (RL) to achieve efficient and robust reasoning capabilities. Built upon the publicly available Ling-lite model, a 16.8 billion parameter model with 2.75 billion activated parameters, our approach matches the performance of state-of-the-art (SOTA) small-scale reasoning models on challenging benchmarks (e.g., AIME, LiveCodeBench, GPQA-Diamond) while activating only one-third of the parameters required by comparable models. To accomplish this, we introduce a joint training pipeline integrating distillation with RL, revealing undocumented challenges in MoE RL training. First, we identify optimization instability during RL training, and we propose Constrained Contextual Computation Policy Optimization(C3PO), a novel approach that enhances training stability and improves computational throughput via algorithm-system co-design methodology. Second, we empirically demonstrate that selecting distillation checkpoints based on entropy loss for RL training, rather than validation metrics, yields superior performance-efficiency trade-offs in subsequent RL training. Finally, we develop a two-stage training paradigm to harmonize multi-domain data integration, addressing domain conflicts that arise in training with mixed dataset. We will release the model, dataset, and code.

This work focuses on the credit assignment problem in cooperative multi-agent reinforcement learning (MARL). Sharing the global advantage among agents often leads to suboptimal policy updates as it fails to account for the distinct contributions of agents. Although numerous methods consider global or individual contributions for credit assignment, a detailed analysis at the coalition level remains lacking in many approaches. This work analyzes the over-updating problem during multi-agent policy updates from a coalition-level perspective. To address this issue, we propose a credit assignment method called Coalitional Rational Advantage Decomposition (CORA). CORA evaluates coalitional advantages via marginal contributions from all possible coalitions and decomposes advantages using the core solution from cooperative game theory, ensuring coalitional rationality. To reduce computational overhead, CORA employs random coalition sampling. Experiments on matrix games, differential games, and multi-agent collaboration benchmarks demonstrate that CORA outperforms strong baselines, particularly in tasks with multiple local optima. These findings highlight the importance of coalition-aware credit assignment for improving MARL performance.

Action-labeled data for robotics is scarce and expensive, limiting the generalization of learned policies. In contrast, vast amounts of action-free video data are readily available, but translating these observations into effective policies remains a challenge. We introduce AMPLIFY, a novel framework that leverages large-scale video data by encoding visual dynamics into compact, discrete motion tokens derived from keypoint trajectories. Our modular approach separates visual motion prediction from action inference, decoupling the challenges of learning what motion defines a task from how robots can perform it. We train a forward dynamics model on abundant action-free videos and an inverse dynamics model on a limited set of action-labeled examples, allowing for independent scaling. Extensive evaluations demonstrate that the learned dynamics are both accurate, achieving up to 3.7x better MSE and over 2.5x better pixel prediction accuracy compared to prior approaches, and broadly useful. In downstream policy learning, our dynamics predictions enable a 1.2-2.2x improvement in low-data regimes, a 1.4x average improvement by learning from action-free human videos, and the first generalization to LIBERO tasks from zero in-distribution action data. Beyond robotic control, we find the dynamics learned by AMPLIFY to be a versatile latent world model, enhancing video prediction quality. Our results present a novel paradigm leveraging heterogeneous data sources to build efficient, generalizable world models. More information can be found at https://amplify-robotics.github.io/.

Reinforcement learning-based traffic signal control (RL-TSC) has emerged as a promising approach for improving urban mobility. However, its robustness under real-world disruptions such as traffic incidents remains largely underexplored. In this study, we introduce T-REX, an open-source, SUMO-based simulation framework for training and evaluating RL-TSC methods under dynamic, incident scenarios. T-REX models realistic network-level performance considering drivers' probabilistic rerouting, speed adaptation, and contextual lane-changing, enabling the simulation of congestion propagation under incidents. To assess robustness, we propose a suite of metrics that extend beyond conventional traffic efficiency measures. Through extensive experiments across synthetic and real-world networks, we showcase T-REX for the evaluation of several state-of-the-art RL-TSC methods under multiple real-world deployment paradigms. Our findings show that while independent value-based and decentralized pressure-based methods offer fast convergence and generalization in stable traffic conditions and homogeneous networks, their performance degrades sharply under incident-driven distribution shifts. In contrast, hierarchical coordination methods tend to offer more stable and adaptable performance in large-scale, irregular networks, benefiting from their structured decision-making architecture. However, this comes with the trade-off of slower convergence and higher training complexity. These findings highlight the need for robustness-aware design and evaluation in RL-TSC research. T-REX contributes to this effort by providing an open, standardized and reproducible platform for benchmarking RL methods under dynamic and disruptive traffic scenarios.

Algorithm Distillation (AD) was recently proposed as a new approach to perform In-Context Reinforcement Learning (ICRL) by modeling across-episodic training histories autoregressively with a causal transformer model. However, due to practical limitations induced by the attention mechanism, experiments were bottlenecked by the transformer's quadratic complexity and limited to simple discrete environments with short time horizons. In this work, we propose leveraging the recently proposed Selective Structured State Space Sequence (S6) models, which achieved state-of-the-art (SOTA) performance on long-range sequence modeling while scaling linearly in sequence length. Through four complex and continuous Meta Reinforcement Learning environments, we demonstrate the overall superiority of Mamba, a model built with S6 layers, over a transformer model for AD. Additionally, we show that scaling AD to very long contexts can improve ICRL performance and make it competitive even with a SOTA online meta RL baseline.

-- Quadrotor Morpho-Transition, or the act of transitioning from air to ground through mid-air transformation, involves complex aerodynamic interactions and a need to operate near actuator saturation, complicating controller design. In recent work, morpho-transition has been studied from a model-based control perspective, but these approaches remain limited due to unmodeled dynamics and the requirement for planning through contacts. Here, we train an end-to-end Reinforcement Learning (RL) controller to learn a morpho-transition policy and demonstrate successful transfer to hardware. We find that the RL control policy achieves agile landing, but only transfers to hardware if motor dynamics and observation delays are taken into account. On the other hand, a baseline MPC controller transfers out-of-the-box without knowledge of the actuator dynamics and delays, at the cost of reduced recovery from disturbances in the event of unknown actuator failures. Our work opens the way for more robust control of agile in-flight quadrotor maneuvers that require mid-air transformation. Ground aerial robotic systems are ideally poised to increase the reliability and scope of autonomous robotic missions.

The classical Job Shop Scheduling Problem (JSSP) focuses on optimizing makespan under deterministic constraints. Real-world production environments introduce additional complexities that cause traditional scheduling approaches to be less effective. Reinforcement learning (RL) holds potential in addressing these challenges, as it allows agents to learn adaptive scheduling strategies. However, there is a lack of a comprehensive, general-purpose frameworks for effectively training and evaluating RL agents under real-world constraints. To address this gap, we propose a modular framework that extends classical JSSP formulations by incorporating key real-world constraints inherent to the shopfloor, including transport logistics, buffer management, machine breakdowns, setup times, and stochastic processing conditions, while also supporting multi-objective optimization. The framework is a customizable solution that offers flexibility in defining problem instances and configuring simulation parameters, enabling adaptation to diverse production scenarios. A standardized interface ensures compatibility with various RL approaches, providing a robust environment for training RL agents and facilitating the standardized comparison of different scheduling methods under dynamic and uncertain conditions. We release JobShopLab as an open-source tool for both research and industrial applications, accessible at: https://github.com/proto-lab-ro/jobshoplab

Preference-based Reinforcement Learning (PbRL) methods provide a solution to avoid reward engineering by learning reward models based on human preferences. However, poor feedback- and sample- efficiency still remain the problems that hinder the application of PbRL. In this paper, we present a novel efficient query selection and preference-guided exploration method, called SENIOR, which could select the meaningful and easy-to-comparison behavior segment pairs to improve human feedback-efficiency and accelerate policy learning with the designed preference-guided intrinsic rewards. Our key idea is twofold: (1) We designed a Motion-Distinction-based Selection scheme (MDS). It selects segment pairs with apparent motion and different directions through kernel density estimation of states, which is more task-related and easy for human preference labeling; (2) We proposed a novel preference-guided exploration method (PGE). It encourages the exploration towards the states with high preference and low visits and continuously guides the agent achieving the valuable samples. The synergy between the two mechanisms could significantly accelerate the progress of reward and policy learning. Our experiments show that SENIOR outperforms other five existing methods in both human feedback-efficiency and policy convergence speed on six complex robot manipulation tasks from simulation and four real-worlds.

Imitation learning has driven the development of generalist policies capable of autonomously solving multiple tasks. However, when a pretrained policy makes errors during deployment, there are limited mechanisms for users to correct its behavior. While collecting additional data for finetuning can address such issues, doing so for each downstream use case is inefficient at deployment. My research proposes an alternative: keeping pretrained policies frozen as a fixed skill repertoire while allowing user interactions to guide behavior generation toward user preferences at inference time. By making pretrained policies steerable, users can help correct policy errors when the model struggles to generalize-without needing to finetune the policy. Specifically, I propose (1) inference-time steering, which leverages user interactions to switch between discrete skills, and (2) task and motion imitation, which enables user interactions to edit continuous motions while satisfying task constraints defined by discrete symbolic plans. These frameworks correct misaligned policy predictions without requiring additional training, maximizing the utility of pretrained models while achieving inference-time user objectives.