We investigate reinforcement learning (RL) for privileged planning in autonomous driving. State-of-the-art approaches for this task are rule-based, but these methods do not scale to the long tail. RL, on the other hand, is scalable and does not suffer from compounding errors like imitation learning. Contemporary RL approaches for driving use complex shaped rewards that sum multiple individual rewards, \eg~progress, position, or orientation rewards. We show that PPO fails to optimize a popular version of these rewards when the mini-batch size is increased, which limits the scalability of these approaches. Instead, we propose a new reward design based primarily on optimizing a single intuitive reward term: route completion. Infractions are penalized by terminating the episode or multiplicatively reducing route completion. We find that PPO scales well with higher mini-batch sizes when trained with our simple reward, even improving performance. Training with large mini-batch sizes enables efficient scaling via distributed data parallelism. We scale PPO to 300M samples in CARLA and 500M samples in nuPlan with a single 8-GPU node. The resulting model achieves 64 DS on the CARLA longest6 v2 benchmark, outperforming other RL methods with more complex rewards by a large margin. Requiring only minimal adaptations from its use in CARLA, the same method is the best learning-based approach on nuPlan. It scores 91.3 in non-reactive and 90.6 in reactive traffic on the Val14 benchmark while being an order of magnitude faster than prior work.
The job shop scheduling problem is an NP-hard combinatorial optimization problem relevant to manufacturing and timetabling. Traditional approaches use priority dispatching rules based on simple heuristics. Recent work has attempted to replace these with deep learning models, particularly graph neural networks (GNNs), that learn to assign priorities from data. However, training such models requires customizing numerous factors: graph representation, node features, action space, and reward functions. The lack of modular libraries for experimentation makes this research time-consuming. This work introduces JobShopLib, a modular library that allows customizing these factors and creating new components with its reinforcement learning environment. We trained several dispatchers through imitation learning to demonstrate the environment's utility. One model outperformed various graph-based dispatchers using only individual operation features, highlighting the importance of feature customization. Our GNN model achieved near state-of-the-art results on large-scale problems. These results suggest significant room for improvement in developing such models. JobShopLib provides the necessary tools for future experimentation.
Interactive Imitation Learning (IIL) allows agents to acquire desired behaviors through human interventions, but current methods impose high cognitive demands on human supervisors. We propose the Adaptive Intervention Mechanism (AIM), a novel robot-gated IIL algorithm that learns an adaptive criterion for requesting human demonstrations. AIM utilizes a proxy Q-function to mimic the human intervention rule and adjusts intervention requests based on the alignment between agent and human actions. By assigning high Q-values when the agent deviates from the expert and decreasing these values as the agent becomes proficient, the proxy Q-function enables the agent to assess the real-time alignment with the expert and request assistance when needed. Our expert-in-the-loop experiments reveal that AIM significantly reduces expert monitoring efforts in both continuous and discrete control tasks. Compared to the uncertainty-based baseline Thrifty-DAgger, our method achieves a 40% improvement in terms of human take-over cost and learning efficiency. Furthermore, AIM effectively identifies safety-critical states for expert assistance, thereby collecting higher-quality expert demonstrations and reducing overall expert data and environment interactions needed. Code and demo video are available at https://github.com/metadriverse/AIM.
The fundamental limitation of the behavioral cloning (BC) approach to imitation learning is that it only teaches an agent what the expert did at states the expert visited. This means that when a BC agent makes a mistake which takes them out of the support of the demonstrations, they often don't know how to recover from it. In this sense, BC is akin to giving the agent the fish -- giving them dense supervision across a narrow set of states -- rather than teaching them to fish: to be able to reason independently about achieving the expert's outcome even when faced with unseen situations at test-time. In response, we explore learning to search (L2S) from expert demonstrations, i.e. learning the components required to, at test time, plan to match expert outcomes, even after making a mistake. These include (1) a world model and (2) a reward model. We carefully ablate the set of algorithmic and design decisions required to combine these and other components for stable and sample/interaction-efficient learning of recovery behavior without additional human corrections. Across a dozen visual manipulation tasks from three benchmarks, our approach $\texttt{SAILOR}$ consistently out-performs state-of-the-art Diffusion Policies trained via BC on the same data. Furthermore, scaling up the amount of demonstrations used for BC by 5-10$\times$ still leaves a performance gap. We find that $\texttt{SAILOR}$ can identify nuanced failures and is robust to reward hacking. Our code is available at https://github.com/arnavkj1995/SAILOR .
Empowering large language models (LLMs) with effective tool utilization capabilities is crucial for enabling AI agents to solve complex problems. However, current models face two major limitations: (1) unreliable tool planning and invocation due to low-quality instruction datasets (e.g., widespread hallucinated API calls), and (2) weak tool reflection abilities (over 90% of errors cannot be corrected) resulting from static imitation learning. To address these critical limitations, we propose Tool-MVR, a novel Tool-Augmented LLM that achieves comprehensive System 2 reasoning through two key innovations. Specifically, we first introduce Multi-Agent Meta-Verification (MAMV), a systematic pipeline that rigorously validates APIs, queries, and reasoning trajectories to construct ToolBench-V, a new high-quality instruction dataset that addresses the limitation of unreliable tool planning and invocation. Second, we propose Exploration-based Reflection Learning (EXPLORE), which enhances tool reflection capabilities by leveraging tool feedback through a dynamic "Error -> Reflection -> Correction" learning paradigm, resulting in our reflection dataset ToolBench-R and addressing the critical weakness in tool reflection. Finally, we obtain Tool-MVR by finetuning open-source LLMs (e.g., Qwen-7B) on both ToolBench-V and ToolBench-R. Our experiments demonstrate that Tool-MVR achieves state-of-the-art performance on StableToolBench, surpassing both ToolLLM (by 23.9%) and GPT-4 (by 15.3%) while reducing API calls by 31.4%, with strong generalization capabilities across unseen tools and scenarios. Additionally, on our proposed RefineToolBench, the first benchmark specifically designed to evaluate tool reflection capabilities, Tool-MVR achieves a 58.9% error correction rate, significantly outperforming ToolLLM's 9.1%.
Autonomous driving promises significant advancements in mobility, road safety and traffic efficiency, yet reinforcement learning and imitation learning face safe-exploration and distribution-shift challenges. Although human-AI collaboration alleviates these issues, it often relies heavily on extensive human intervention, which increases costs and reduces efficiency. This paper develops a confidence-guided human-AI collaboration (C-HAC) strategy to overcome these limitations. First, C-HAC employs a distributional proxy value propagation method within the distributional soft actor-critic (DSAC) framework. By leveraging return distributions to represent human intentions C-HAC achieves rapid and stable learning of human-guided policies with minimal human interaction. Subsequently, a shared control mechanism is activated to integrate the learned human-guided policy with a self-learning policy that maximizes cumulative rewards. This enables the agent to explore independently and continuously enhance its performance beyond human guidance. Finally, a policy confidence evaluation algorithm capitalizes on DSAC's return distribution networks to facilitate dynamic switching between human-guided and self-learning policies via a confidence-based intervention function. This ensures the agent can pursue optimal policies while maintaining safety and performance guarantees. Extensive experiments across diverse driving scenarios reveal that C-HAC significantly outperforms conventional methods in terms of safety, efficiency, and overall performance, achieving state-of-the-art results. The effectiveness of the proposed method is further validated through real-world road tests in complex traffic conditions. The videos and code are available at: https://github.com/lzqw/C-HAC.
Modern reinforcement learning (RL) algorithms have found success by using powerful probabilistic models, such as transformers, energy-based models, and diffusion/flow-based models. To this end, RL researchers often choose to pay the price of accommodating these models into their algorithms -- diffusion models are expressive, but are computationally intensive due to their reliance on solving differential equations, while autoregressive transformer models are scalable but typically require learning discrete representations. Normalizing flows (NFs), by contrast, seem to provide an appealing alternative, as they enable likelihoods and sampling without solving differential equations or autoregressive architectures. However, their potential in RL has received limited attention, partly due to the prevailing belief that normalizing flows lack sufficient expressivity. We show that this is not the case. Building on recent work in NFs, we propose a single NF architecture which integrates seamlessly into RL algorithms, serving as a policy, Q-function, and occupancy measure. Our approach leads to much simpler algorithms, and achieves higher performance in imitation learning, offline, goal conditioned RL and unsupervised RL.
We study the problem of offline imitation learning in Markov decision processes (MDPs), where the goal is to learn a well-performing policy given a dataset of state-action pairs generated by an expert policy. Complementing a recent line of work on this topic that assumes the expert belongs to a tractable class of known policies, we approach this problem from a new angle and leverage a different type of structural assumption about the environment. Specifically, for the class of linear $Q^\pi$-realizable MDPs, we introduce a new algorithm called saddle-point offline imitation learning (\SPOIL), which is guaranteed to match the performance of any expert up to an additive error $\varepsilon$ with access to $\mathcal{O}(\varepsilon^{-2})$ samples. Moreover, we extend this result to possibly non-linear $Q^\pi$-realizable MDPs at the cost of a worse sample complexity of order $\mathcal{O}(\varepsilon^{-4})$. Finally, our analysis suggests a new loss function for training critic networks from expert data in deep imitation learning. Empirical evaluations on standard benchmarks demonstrate that the neural net implementation of \SPOIL is superior to behavior cloning and competitive with state-of-the-art algorithms.
Reinforcement learning (RL) is increasingly applied to real-world problems involving complex and structured decisions, such as routing, scheduling, and assortment planning. These settings challenge standard RL algorithms, which struggle to scale, generalize, and exploit structure in the presence of combinatorial action spaces. We propose Structured Reinforcement Learning (SRL), a novel actor-critic framework that embeds combinatorial optimization layers into the actor neural network. We enable end-to-end learning of the actor via Fenchel-Young losses and provide a geometric interpretation of SRL as a primal-dual algorithm in the dual of the moment polytope. Across six environments with exogenous and endogenous uncertainty, SRL matches or surpasses the performance of unstructured RL and imitation learning on static tasks and improves over these baselines by up to 92% on dynamic problems, with improved stability and convergence speed.
Due to the powerful vision-language reasoning and generalization abilities, multimodal large language models (MLLMs) have garnered significant attention in the field of end-to-end (E2E) autonomous driving. However, their application to closed-loop systems remains underexplored, and current MLLM-based methods have not shown clear superiority to mainstream E2E imitation learning approaches. In this work, we propose ReasonPlan, a novel MLLM fine-tuning framework designed for closed-loop driving through holistic reasoning with a self-supervised Next Scene Prediction task and supervised Decision Chain-of-Thought process. This dual mechanism encourages the model to align visual representations with actionable driving context, while promoting interpretable and causally grounded decision making. We curate a planning-oriented decision reasoning dataset, namely PDR, comprising 210k diverse and high-quality samples. Our method outperforms the mainstream E2E imitation learning method by a large margin of 19% L2 and 16.1 driving score on Bench2Drive benchmark. Furthermore, ReasonPlan demonstrates strong zero-shot generalization on unseen DOS benchmark, highlighting its adaptability in handling zero-shot corner cases. Code and dataset will be found in https://github.com/Liuxueyi/ReasonPlan.
Reinforcement Learning (RL) in partially observable environments poses significant challenges due to the complexity of learning under uncertainty. While additional information, such as that available in simulations, can enhance training, effectively leveraging it remains an open problem. To address this, we introduce Guided Policy Optimization (GPO), a framework that co-trains a guider and a learner. The guider takes advantage of privileged information while ensuring alignment with the learner's policy that is primarily trained via imitation learning. We theoretically demonstrate that this learning scheme achieves optimality comparable to direct RL, thereby overcoming key limitations inherent in existing approaches. Empirical evaluations show strong performance of GPO across various tasks, including continuous control with partial observability and noise, and memory-based challenges, significantly outperforming existing methods.
Generalist imitation learning policies trained on large datasets show great promise for solving diverse manipulation tasks. However, to ensure generalization to different conditions, policies need to be trained with data collected across a large set of environmental factor variations (e.g., camera pose, table height, distractors) $-$ a prohibitively expensive undertaking, if done exhaustively. We introduce a principled method for deciding what data to collect and how much to collect for each factor by constructing factored scaling curves (FSC), which quantify how policy performance varies as data scales along individual or paired factors. These curves enable targeted data acquisition for the most influential factor combinations within a given budget. We evaluate the proposed method through extensive simulated and real-world experiments, across both training-from-scratch and fine-tuning settings, and show that it boosts success rates in real-world tasks in new environments by up to 26% over existing data-collection strategies. We further demonstrate how factored scaling curves can effectively guide data collection using an offline metric, without requiring real-world evaluation at scale.
This work focuses on the design of a deep learning-based autonomous driving system deployed and tested on the real-world MIT Racecar to assess its effectiveness in driving scenarios. The Deep Neural Network (DNN) translates raw image inputs into real-time steering commands in an end-to-end learning fashion, following the imitation learning framework. The key design challenge is to ensure that DNN predictions are accurate and fast enough, at a high sampling frequency, and result in smooth vehicle operation under different operating conditions. In this study, we design and compare various DNNs, to identify the most effective approach for real-time autonomous driving. In designing the DNNs, we adopted an incremental design approach that involved enhancing the model capacity and dataset to address the challenges of real-world driving scenarios. We designed a PD system, CNN, CNN-LSTM, and CNN-NODE, and evaluated their performance on the real-world MIT Racecar. While the PD system handled basic lane following, it struggled with sharp turns and lighting variations. The CNN improved steering but lacked temporal awareness, which the CNN-LSTM addressed as it resulted in smooth driving performance. The CNN-NODE performed similarly to the CNN-LSTM in handling driving dynamics, yet with slightly better driving performance. The findings of this research highlight the importance of iterative design processes in developing robust DNNs for autonomous driving applications. The experimental video is available at https://www.youtube.com/watch?v=FNNYgU--iaY.
Understanding action correspondence between humans and robots is essential for evaluating alignment in decision-making, particularly in human-robot collaboration and imitation learning within unstructured environments. We propose a multimodal demonstration learning framework that explicitly models human demonstrations from RGB video with robot demonstrations in voxelized RGB-D space. Focusing on the "pick and place" task from the RH20T dataset, we utilize data from 5 users across 10 diverse scenes. Our approach combines ResNet-based visual encoding for human intention modeling and a Perceiver Transformer for voxel-based robot action prediction. After 2000 training epochs, the human model reaches 71.67% accuracy, and the robot model achieves 71.8% accuracy, demonstrating the framework's potential for aligning complex, multimodal human and robot behaviors in manipulation tasks.
Language models rely on semantic priors to perform in-context learning, which leads to poor performance on tasks involving inductive reasoning. Instruction-tuning methods based on imitation learning can superficially enhance the in-context learning performance of language models, but they often fail to improve the model's understanding of the underlying rules that connect inputs and outputs in few-shot demonstrations. We propose ReDis, a reasoning distillation technique designed to improve the inductive reasoning capabilities of language models. Through a careful combination of data augmentation, filtering, supervised fine-tuning, and alignment, ReDis achieves significant performance improvements across a diverse range of tasks, including 1D-ARC, List Function, ACRE, and MiniSCAN. Experiments on three language model backbones show that ReDis outperforms equivalent few-shot prompting baselines across all tasks and even surpasses the teacher model, GPT-4o, in some cases. ReDis, based on the LLaMA-3 backbone, achieves relative improvements of 23.2%, 2.8%, and 66.6% over GPT-4o on 1D-ARC, ACRE, and MiniSCAN, respectively, within a similar hypothesis search space. The code, dataset, and model checkpoints will be made available at https://github.com/NafisSadeq/reasoning-distillation.git.
Recent advancements in diffusion-based imitation learning, which show impressive performance in modeling multimodal distributions and training stability, have led to substantial progress in various robot learning tasks. In visual navigation, previous diffusion-based policies typically generate action sequences by initiating from denoising Gaussian noise. However, the target action distribution often diverges significantly from Gaussian noise, leading to redundant denoising steps and increased learning complexity. Additionally, the sparsity of effective action distributions makes it challenging for the policy to generate accurate actions without guidance. To address these issues, we propose a novel, unified visual navigation framework leveraging the denoising diffusion bridge models named NaviBridger. This approach enables action generation by initiating from any informative prior actions, enhancing guidance and efficiency in the denoising process. We explore how diffusion bridges can enhance imitation learning in visual navigation tasks and further examine three source policies for generating prior actions. Extensive experiments in both simulated and real-world indoor and outdoor scenarios demonstrate that NaviBridger accelerates policy inference and outperforms the baselines in generating target action sequences. Code is available at https://github.com/hren20/NaiviBridger.
Assistance games are a promising alternative to reinforcement learning from human feedback (RLHF) for training AI assistants. Assistance games resolve key drawbacks of RLHF, such as incentives for deceptive behavior, by explicitly modeling the interaction between assistant and user as a two-player game where the assistant cannot observe their shared goal. Despite their potential, assistance games have only been explored in simple settings. Scaling them to more complex environments is difficult because it requires both solving intractable decision-making problems under uncertainty and accurately modeling human users' behavior. We present the first scalable approach to solving assistance games and apply it to a new, challenging Minecraft-based assistance game with over $10^{400}$ possible goals. Our approach, AssistanceZero, extends AlphaZero with a neural network that predicts human actions and rewards, enabling it to plan under uncertainty. We show that AssistanceZero outperforms model-free RL algorithms and imitation learning in the Minecraft-based assistance game. In a human study, our AssistanceZero-trained assistant significantly reduces the number of actions participants take to complete building tasks in Minecraft. Our results suggest that assistance games are a tractable framework for training effective AI assistants in complex environments. Our code and models are available at https://github.com/cassidylaidlaw/minecraft-building-assistance-game.
Imitation learning based planning tasks on the nuPlan dataset have gained great interest due to their potential to generate human-like driving behaviors. However, open-loop training on the nuPlan dataset tends to cause causal confusion during closed-loop testing, and the dataset also presents a long-tail distribution of scenarios. These issues introduce challenges for imitation learning. To tackle these problems, we introduce CAFE-AD, a Cross-Scenario Adaptive Feature Enhancement for Trajectory Planning in Autonomous Driving method, designed to enhance feature representation across various scenario types. We develop an adaptive feature pruning module that ranks feature importance to capture the most relevant information while reducing the interference of noisy information during training. Moreover, we propose a cross-scenario feature interpolation module that enhances scenario information to introduce diversity, enabling the network to alleviate over-fitting in dominant scenarios. We evaluate our method CAFE-AD on the challenging public nuPlan Test14-Hard closed-loop simulation benchmark. The results demonstrate that CAFE-AD outperforms state-of-the-art methods including rule-based and hybrid planners, and exhibits the potential in mitigating the impact of long-tail distribution within the dataset. Additionally, we further validate its effectiveness in real-world environments. The code and models will be made available at https://github.com/AlniyatRui/CAFE-AD.
Many recent advances in robotic manipulation have come through imitation learning, yet these rely largely on mimicking a particularly hard-to-acquire form of demonstrations: those collected on the same robot in the same room with the same objects as the trained policy must handle at test time. In contrast, large pre-recorded human video datasets demonstrating manipulation skills in-the-wild already exist, which contain valuable information for robots. Is it possible to distill a repository of useful robotic skill policies out of such data without any additional requirements on robot-specific demonstrations or exploration? We present the first such system ZeroMimic, that generates immediately deployable image goal-conditioned skill policies for several common categories of manipulation tasks (opening, closing, pouring, pick&place, cutting, and stirring) each capable of acting upon diverse objects and across diverse unseen task setups. ZeroMimic is carefully designed to exploit recent advances in semantic and geometric visual understanding of human videos, together with modern grasp affordance detectors and imitation policy classes. After training ZeroMimic on the popular EpicKitchens dataset of ego-centric human videos, we evaluate its out-of-the-box performance in varied real-world and simulated kitchen settings with two different robot embodiments, demonstrating its impressive abilities to handle these varied tasks. To enable plug-and-play reuse of ZeroMimic policies on other task setups and robots, we release software and policy checkpoints of our skill policies.
Recent advances in deep thinking models have demonstrated remarkable reasoning capabilities on mathematical and coding tasks. However, their effectiveness in embodied domains which require continuous interaction with environments through image action interleaved trajectories remains largely -unexplored. We present Embodied Reasoner, a model that extends o1 style reasoning to interactive embodied search tasks. Unlike mathematical reasoning that relies primarily on logical deduction, embodied scenarios demand spatial understanding, temporal reasoning, and ongoing self-reflection based on interaction history. To address these challenges, we synthesize 9.3k coherent Observation-Thought-Action trajectories containing 64k interactive images and 90k diverse thinking processes (analysis, spatial reasoning, reflection, planning, and verification). We develop a three-stage training pipeline that progressively enhances the model's capabilities through imitation learning, self-exploration via rejection sampling, and self-correction through reflection tuning. The evaluation shows that our model significantly outperforms those advanced visual reasoning models, e.g., it exceeds OpenAI o1, o3-mini, and Claude-3.7 by +9\%, 24\%, and +13\%. Analysis reveals our model exhibits fewer repeated searches and logical inconsistencies, with particular advantages in complex long-horizon tasks. Real-world environments also show our superiority while exhibiting fewer repeated searches and logical inconsistency cases.