In the supplementary, we first discuss the experimental details and hyperparameters in Section A. Section B, and further present the visualization in RLBench in Section C. Finally, we discuss how to MLP with 512 hidden units and Mish activations. The probability ϵ of random actions is set to 0. 03 in Stochastic Environments. So the sampled trajectories still lead to the collision. Figure 1 illustrates a problematic sampled trajectory after execution. We further evaluate the performance with different replanning steps in Table 1.

In the supplementary, we first discuss the experimental details and hyperparameters in Section A. Section B, and further present the visualization in RLBench in Section C. Finally, we discuss how to MLP with 512 hidden units and Mish activations. The probability ϵ of random actions is set to 0. 03 in Stochastic Environments. So the sampled trajectories still lead to the collision. Figure 1 illustrates a problematic sampled trajectory after execution. We further evaluate the performance with different replanning steps in Table 1.

Large language models (LLMs) are increasingly explored in robot manipulation, but many existing methods struggle to adapt to new environments. Many systems require either environment-specific policy training or depend on fixed prompts and single-shot code generation, leading to limited transferability and manual re-tuning. We introduce Memory Transfer Planning (MTP), a framework that leverages successful control-code examples from different environments as procedural knowledge, using them as in-context guidance for LLM-driven planning. Specifically, MTP (i) generates an initial plan and code using LLMs, (ii) retrieves relevant successful examples from a code memory, and (iii) contextually adapts the retrieved code to the target setting for re-planning without updating model parameters. We evaluate MTP on RLBench, CALVIN, and a physical robot, demonstrating effectiveness beyond simulation. Across these settings, MTP consistently improved success rate and adaptability compared with fixed-prompt code generation, naive retrieval, and memory-free re-planning. Furthermore, in hardware experiments, leveraging a memory constructed in simulation proved effective. MTP provides a practical approach that exploits procedural knowledge to realize robust LLM-based planning across diverse robotic manipulation scenarios, enhancing adaptability to novel environments and bridging simulation and real-world deployment.

Effective robotic manipulation relies on a precise understanding of 3D scene geometry, and one of the most straightforward ways to acquire such geometry is through multi-view observations. Motivated by this, we present GP3 -- a 3D geometry-aware robotic manipulation policy that leverages multi-view input. GP3 employs a spatial encoder to infer dense spatial features from RGB observations, which enable the estimation of depth and camera parameters, leading to a compact yet expressive 3D scene representation tailored for manipulation. This representation is fused with language instructions and translated into continuous actions via a lightweight policy head. Comprehensive experiments demonstrate that GP3 consistently outperforms state-of-the-art methods on simulated benchmarks. Furthermore, GP3 transfers effectively to real-world robots without depth sensors or pre-mapped environments, requiring only minimal fine-tuning. These results highlight GP3 as a practical, sensor-agnostic solution for geometry-aware robotic manipulation.

Despite significant advancements in robotic manipulation, achieving consistent and stable grasping remains a fundamental challenge, often limiting the successful execution of complex tasks. Our analysis reveals that even state-of-the-art policy models frequently exhibit unstable grasping behaviors, leading to failure cases that create bottlenecks in real-world robotic applications. To address these challenges, we introduce GraspCorrect, a plug-and-play module designed to enhance grasp performance through vision-language model-guided feedback. GraspCorrect employs an iterative visual question-answering framework with two key components: grasp-guided prompting, which incorporates task-specific constraints, and object-aware sampling, which ensures the selection of physically feasible grasp candidates. By iteratively generating intermediate visual goals and translating them into joint-level actions, GraspCorrect significantly improves grasp stability and consistently enhances task success rates across existing policy models in the RLBench and CALVIN datasets.

Reinforcement learning (RL) for continuous control often requires large amounts of online interaction data. Value-based RL methods can mitigate this burden by offering relatively high sample efficiency. Some studies further enhance sample efficiency by incorporating offline demonstration data to "kick-start" training, achieving promising results in continuous control. However, they typically compute the Q-function independently for each action dimension, neglecting interdependencies and making it harder to identify optimal actions when learning from suboptimal data, such as non-expert demonstration and online-collected data during the training process. To address these issues, we propose Auto-Regressive Soft Q-learning (ARSQ), a value-based RL algorithm that models Q-values in a coarse-to-fine, auto-regressive manner. First, ARSQ decomposes the continuous action space into discrete spaces in a coarse-to-fine hierarchy, enhancing sample efficiency for fine-grained continuous control tasks. Next, it auto-regressively predicts dimensional action advantages within each decision step, enabling more effective decision-making in continuous control tasks. We evaluate ARSQ on two continuous control benchmarks, RLBench and D4RL, integrating demonstration data into online training. On D4RL, which includes non-expert demonstrations, ARSQ achieves an average $1.62\times$ performance improvement over SOTA value-based baseline. On RLBench, which incorporates expert demonstrations, ARSQ surpasses various baselines, demonstrating its effectiveness in learning from suboptimal online-collected data.

Efficient control in long-horizon robotic manipulation is challenging due to complex representation and policy learning requirements. Model-based visual reinforcement learning (RL) has shown great potential in addressing these challenges but still faces notable limitations, particularly in handling sparse rewards and complex visual features in long-horizon environments. To address these limitations, we propose the Recognize-Sense-Plan-Act (RSPA) pipeline for long-horizon tasks and further introduce RoboHorizon, an LLM-assisted multi-view world model tailored for long-horizon robotic manipulation. In RoboHorizon, pre-trained LLMs generate dense reward structures for multi-stage sub-tasks based on task language instructions, enabling robots to better recognize long-horizon tasks. Keyframe discovery is then integrated into the multi-view masked autoencoder (MAE) architecture to enhance the robot's ability to sense critical task sequences, strengthening its multi-stage perception of long-horizon processes. Leveraging these dense rewards and multi-view representations, a robotic world model is constructed to efficiently plan long-horizon tasks, enabling the robot to reliably act through RL algorithms. Experiments on two representative benchmarks, RLBench and FurnitureBench, show that RoboHorizon outperforms state-of-the-art visual model-based RL methods, achieving a 23.35% improvement in task success rates on RLBench's 4 short-horizon tasks and a 29.23% improvement on 6 long-horizon tasks from RLBench and 3 furniture assembly tasks from FurnitureBench.

Training reinforcement learning (RL) agents on robotic tasks typically requires a large number of training samples. This is because training data often consists of noisy trajectories, whether from exploration or human-collected demonstrations, making it difficult to learn value functions that understand the effect of taking each action. On the other hand, recent behavior-cloning (BC) approaches have shown that predicting a sequence of actions enables policies to effectively approximate noisy, multi-modal distributions of expert demonstrations. Can we use a similar idea for improving RL on robotic tasks? In this paper, we introduce a novel RL algorithm that learns a critic network that outputs Q-values over a sequence of actions. By explicitly training the value functions to learn the consequence of executing a series of current and future actions, our algorithm allows for learning useful value functions from noisy trajectories. We study our algorithm across various setups with sparse and dense rewards, and with or without demonstrations, spanning mobile bi-manual manipulation, whole-body control, and tabletop manipulation tasks from BiGym, HumanoidBench, and RLBench. We find that, by learning the critic network with action sequences, our algorithm outperforms various RL and BC baselines, in particular on challenging humanoid control tasks.

Designing a universal policy architecture that performs well across diverse robots and task configurations remains a key challenge. In this work, we address this by representing robot actions as sequential data and generating actions through autoregressive sequence modeling. Existing autoregressive architectures generate end-effector waypoints sequentially as word tokens in language modeling, which are limited to low-frequency control tasks. Unlike language, robot actions are heterogeneous and often include continuous values -- such as joint positions, 2D pixel coordinates, and end-effector poses -- which are not easily suited for language-based modeling. Based on this insight, we introduce a straightforward enhancement: we extend causal transformers' single-token prediction to support predicting a variable number of tokens in a single step through our Chunking Causal Transformer (CCT). This enhancement enables robust performance across diverse tasks of various control frequencies, greater efficiency by having fewer autoregression steps, and lead to a hybrid action sequence design by mixing different types of actions and using a different chunk size for each action type. Based on CCT, we propose the Autoregressive Policy (ARP) architecture, which solves manipulation tasks by generating hybrid action sequences. We evaluate ARP across diverse robotic manipulation environments, including Push-T, ALOHA, and RLBench, and show that ARP, as a universal architecture, outperforms the environment-specific state-of-the-art in all tested benchmarks, while being more efficient in computation and parameter sizes. Videos of our real robot demonstrations, all source code and the pretrained models of ARP can be found at http://github.com/mlzxy/arp.