VLA 研究日報VLA 研究日报

Generative policies based on diffusion and flow matching achieve strong performance in robotic manipulation by modeling multi-modal human demonstrations. However, their reliance on iterative Ordinary Differential Equation (ODE) integration introduces substantial latency, limiting high-frequency closed-loop control. Recent single-step acceleration methods alleviate this overhead but often exhibit distributional collapse, producing averaged trajectories that fail to execute coherent manipulation strategies. We propose a framework that distills a Conditional Flow Matching (CFM) expert into a fast single-step student via Implicit Maximum Likelihood Estimation (IMLE). A bi-directional Chamfer distance provides a set-level objective that promotes both mode coverage and fidelity, enabling preservation of the teacher multi-modal action distribution in a single forward pass. A unified perception encoder further integrates multi-view RGB, depth, point clouds, and proprioception into a geometry-aware representation. The resulting high-frequency control supports real-time receding-horizon re-planning and improved robustness under dynamic disturbances.

Deploying Vision-Language-Action (VLA) models in real-world robotics exposes a core multi-task learning challenge: reconciling task interference in multi-task robotic learning. When multiple tasks are jointly fine-tuned in a single stage, gradients from different tasks can conflict, causing negative transfer and reducing per-task performance. Yet maintaining a separate full checkpoint per task is often storage- and deployment-prohibitive. To address this dilemma, we present CORAL, a backbone- and embodiment-agnostic framework designed primarily to mitigate multi-task interference while remaining naturally extensible to a continuous stream of new tasks. CORAL freezes a single pre-trained VLA backbone and attaches one lightweight Low-Rank Adaptation (LoRA) expert per task; at runtime, a dynamic inference engine (the CORAL Manager) routes language instructions to the appropriate expert and swaps experts on the fly with zero inference overhead. This strict parameter isolation avoids complex gating networks and prevents parameter-level cross-task interference by construction; as an added capability, it also enables sequentially introducing new tasks without parameter overwriting caused by catastrophic forgetting. We validate CORAL on a real-world Galaxea R1 dual-arm mobile manipulator and three simulation benchmarks (LIBERO, WidowX, Google Robot), where CORAL overcomes fine-grained instructional ambiguity and substantially outperforms joint training, yielding a practical and scalable system for lifelong multi-task robot learning. Website: https://frontierrobo.github.io/CORAL

[Huawei]

Measurement of task progress through explicit, actionable milestones is critical for robust robotic manipulation. This progress awareness enables a model to ground its current task status, anticipate verifiable intermediate states, and detect and recover from failures when progress stalls. To embody this capability, we introduce See, Plan, Rewind (SPR), a progress-aware vision-language-action framework that dynamically grounds language instructions into a sequence of spatial subgoals. SPR operates through a continuous core cycle, Seeing the current state and upcoming milestone, Planning a trajectory towards the next 2D waypoint, and Rewinding to a recoverable state upon failure by monitoring progress against the expected sequence. This closed-loop approach enables robust error correction without requiring additional training data or auxiliary models. Extensive experiments demonstrate the framework's effectiveness, generalization and robustness: SPR outperforms the MolmoAct baseline by 5\% on the LIBERO benchmark. On the challenging LIBERO-Plus benchmark with unseen instructions and initial states, SPR achieves state-of-the-art robustness with the smallest performance drop, surpassing OpenVLA-OFT and UniVLA, demonstrating superior out-of-distribution robustness.

[MBZUAI]

Real-world robotic manipulation in homes and factories demands reliability, efficiency, and robustness that approach or surpass those of skilled human operators. We present RL-100, a real-world reinforcement learning framework built on diffusion visuomotor policies. RL-100 unifies imitation and reinforcement learning under a single clipped PPO surrogate objective applied within the denoising process, yielding conservative and stable improvements across offline and online stages. To meet deployment latency requirements, a lightweight consistency distillation method compresses multi-step diffusion into a one-step controller for high-frequency control. The framework is task-, embodiment-, and representation-agnostic, and supports both single-action and action-chunking control. We evaluate RL-100 on eight diverse real-robot tasks, from dynamic pushing and agile bowling to pouring, cloth folding, unscrewing, multi-stage juicing, and long-horizon box folding. RL-100 attains 100 percent success across evaluated trials, for a total of 1000 out of 1000 episodes, including up to 250 out of 250 consecutive trials on one task. It matches or surpasses expert teleoperators in time to completion. Without retraining, a single policy attains approximately 90 percent zero-shot success under environmental and dynamics shifts, adapts in a few-shot regime to significant task variations (86.7 percent), and remains robust to aggressive human perturbations (about 96 percent). Notably, our juicing robot served random customers continuously for about seven hours without failure when deployed zero-shot in a shopping mall. These results suggest a practical path to deployment-ready robot learning by starting from human priors, aligning training objectives with human-grounded metrics, and reliably extending performance beyond human demonstrations.

Diffusion-based models for robotic control, including vision-language-action (VLA) and vision-action (VA) policies, have demonstrated significant capabilities. Yet their advancement is constrained by the high cost of acquiring large-scale interaction datasets. This work introduces an alternative paradigm for enhancing policy performance without additional model training. Perhaps surprisingly, we demonstrate that the composed policies can exceed the performance of either parent policy. Our contribution is threefold. First, we establish a theoretical foundation showing that the convex composition of distributional scores from multiple diffusion models can yield a superior one-step functional objective compared to any individual score. A Grönwall-type bound is then used to show that this single-step improvement propagates through entire generation trajectories, leading to systemic performance gains. Second, motivated by these results, we propose General Policy Composition (GPC), a training-free method that enhances performance by combining the distributional scores of multiple pre-trained policies via a convex combination and test-time search. GPC is versatile, allowing for the plug-and-play composition of heterogeneous policies, including VA and VLA models, as well as those based on diffusion or flow-matching, irrespective of their input visual modalities. Third, we provide extensive empirical validation. Experiments on Robomimic, PushT, and RoboTwin benchmarks, alongside real-world robotic evaluations, confirm that GPC consistently improves performance and adaptability across a diverse set of tasks. Further analysis of alternative composition operators and weighting strategies offers insights into the mechanisms underlying the success of GPC. These results establish GPC as a simple yet effective method for improving control performance by leveraging existing policies.

Vision-Language-Action models (VLAs) mark a major shift in robot learning. They replace specialized architectures and task-tailored components of expert policies with large-scale data collection and setup-specific fine-tuning. In this machine learning-focused workflow that is centered around models and scalable training, traditional robotics software frameworks become a bottleneck, while robot simulations offer only limited support for transitioning from and to real-world experiments. In this work, we close this gap by introducing Robot Control Stack (RCS), a lean ecosystem designed from the ground up to support research in robot learning with large-scale generalist policies. At its core, RCS features a modular and easily extensible layered architecture with a unified interface for simulated and physical robots, facilitating sim-to-real transfer. Despite its minimal footprint and dependencies, it offers a complete feature set, enabling both real-world experiments and large-scale training in simulation. Our contribution is twofold: First, we introduce the architecture of RCS and explain its design principles. Second, we evaluate its usability and performance along the development cycle of VLA and RL policies. Our experiments also provide an extensive evaluation of Octo, OpenVLA, and Pi Zero on multiple robots and shed light on how simulation data can improve real-world policy performance. Our code, datasets, weights, and videos are available at: https://robotcontrolstack.github.io/

The performance of learned robot visuomotor policies is heavily dependent on the size and quality of the training dataset. Although large-scale robot and human datasets are increasingly available, embodiment gaps and mismatched action spaces make them difficult to leverage. Our main insight is that skills performed across different embodiments produce visual similarities in motions that can be captured using off-the-shelf action representations such as optical flow. Moreover, World Models (WMs) can leverage sub-optimal data since they focus on modeling dynamics. In this work, we aim to improve visuomotor policies in low-data regimes by first pretraining a WM using optical flow as an embodiment-agnostic action representation to leverage accessible or easily collected data from multiple embodiments (robots, humans). Given a small set of demonstrations on a target embodiment, we finetune the WM on this data to better align the WM predictions, train a base policy, and learn a robust value function. Using our finetuned WM and value function, our approach evaluates action candidates from the base policy and selects the best one to improve performance. Our approach, which we term Latent Policy Steering (LPS), improves behavior-cloned policies by 10.6% on average across four Robomimic tasks, even though most of the pretraining data comes from the real world. In the real-world experiments, LPS achieves larger gains: 70% relative improvement with 30-50 target-embodiment demonstrations, and 44% relative improvement with 60-100 demonstrations, compared to a behavior-cloned baseline.

[CMU]

VLA models encode visual observations as 2D patch tokens with no intrinsic geometric structure. We introduce GST-VLA with two contributions. First, the Gaussian Spatial Tokenizer (GST) converts frozen dense depth and frozen semantic patch features into $N_g{=}128$ anisotropic 3D Gaussian primitives, each parameterized by a metric residual mean $μ\in \mathbb{R}^3$, log-scale covariance $\log σ\in \mathbb{R}^3$, and learned opacity $α\in (0,1)$. The covariance eigenstructure encodes local surface orientation, and opacity provides per-primitive geometric confidence, both inaccessible from scalar depth. Spatial attention pooling with learned queries concentrates the fixed token budget on geometrically salient regions rather than distributing uniformly. Second, 3D Depth-Aware Chain-of-Thought (DA-CoT) reasoning supervises four structured intermediate spatial thoughts, covering 3D object grounding, grasp affordance contact geometry, pairwise metric distances, and coarse SE(3) waypoints, as explicit generation targets in the training loss. A cross-attention sublayer at every VLM transformer block provides direct access to the raw 256-primitive Gaussian field during DA-CoT generation. A 300M-parameter flow-matching action expert with mixture-of-experts feedforward sublayers decodes 7-DoF delta action chunks via conditional ODE integration, conditioned on both VLM hidden states and DA-CoT outputs through dual cross-attention. Trained with composite $\mathcal{L}_\mathrm{flow} + \mathcal{L}_\mathrm{CoT} + \mathcal{L}_\mathrm{depth}$ across three progressive stages, GST-VLA achieves 96.4% on LIBERO (+2.0%), and 80.2% on SimplerEnv (+5.4%). Ablations isolate the contribution of each GST component, each DA-CoT thought, and each training stage, confirming independent and synergistic gains concentrated on precision demanding tasks.

Vision-Language-Action (VLA) models are formulated to ground instructions in visual context and generate action sequences for robotic manipulation. Despite recent progress, VLA models still face challenges in learning related and reusable primitives, reducing reliance on large-scale data and complex architectures, and enabling exploration beyond demonstrations. To address these challenges, we propose a novel Neuro-Symbolic Vision-Language-Action (NS-VLA) framework via online reinforcement learning (RL). It introduces a symbolic encoder to embedding vision and language features and extract structured primitives, utilizes a symbolic solver for data-efficient action sequencing, and leverages online RL to optimize generation via expansive exploration. Experiments on robotic manipulation benchmarks demonstrate that NS-VLA outperforms previous methods in both one-shot training and data-perturbed settings, while simultaneously exhibiting superior zero-shot generalizability, high data efficiency and expanded exploration space. Our code is available.

While Vision-Language-Action (VLA) models have demonstrated promising generalization capabilities in robotic manipulation, deploying them on specific and complex downstream tasks still demands effective post-training. In parallel, Human-in-the-Loop (HiL) learning has proven to be a powerful mechanism for refining robot policies. However, extending this paradigm to dexterous manipulation remains challenging: multi-finger control is high-dimensional, contact-intensive, and exhibits execution distributions that differ markedly from standard arm motions, leaving existing dexterous VLA systems limited in reliability and adaptability. We present DexHiL, the first integrated arm-hand human-in-the-loop framework for dexterous VLA models, enabling coordinated interventions over the arm and the dexterous hand within a single system. DexHiL introduces an intervention-aware data sampling strategy that prioritizes corrective segments for post-training, alongside a lightweight teleoperation interface that supports instantaneous human corrections during execution. Real-robot experiments demonstrate that DexHiL serves as an effective post-training framework, yielding a substantial performance leap, outperforming standard offline-only fine-tuning baselines by an average of 25% in success rates across distinct tasks. Project page: https://chenzhongxi-sjtu.github.io/dexhil/

Action-conditioned video models offer a promising path to building general-purpose robot simulators that can improve directly from data. Yet, despite training on large-scale robot datasets, current state-of-the-art video models still struggle to predict physically consistent robot-object interactions that are crucial in robotic manipulation. To close this gap, we present PlayWorld, a simple, scalable, and fully autonomous pipeline for training high-fidelity video world simulators from interaction experience. In contrast to prior approaches that rely on success-biased human demonstrations, PlayWorld is the first system capable of learning entirely from unsupervised robot self-play, enabling naturally scalable data collection while capturing complex, long-tailed physical interactions essential for modeling realistic object dynamics. Experiments across diverse manipulation tasks show that PlayWorld generates high-quality, physically consistent predictions for contact-rich interactions that are not captured by world models trained on human-collected data. We further demonstrate the versatility of PlayWorld in enabling fine-grained failure prediction and policy evaluation, with up to 40% improvements over human-collected data. Finally, we demonstrate how PlayWorld enables reinforcement learning in the world model, improving policy performance by 65% in success rates when deployed in the real world.