New robotics and physical-AI papers, with what each one means for the industry.
Pretrained vision-language-action (VLA) models encode broad manipulation priors, yet adapting them to specific tasks requires balancing limited robot interaction, computational resources, and safety constraints. Noise space learning (NSL) adapts flow-matching-based VLAs by learning a context-conditioned distribution over noise while keeping the pretrained VLA fixed.
Actions specify what a robot should execute, but do not describe what those commands will achieve. Vision-language-action (VLA) policies nevertheless learn primarily by predicting action chunks from images, language, and state.
A user who brings a generalist manipulation policy into their home or office wants it to pick up their objects from their shelves, counters, and tables. Yet even a strong generalist like π_(0.5) achieves 0% pickup success in our captured deployment scenes without adaptation, and collecting real demonstrations in every user's environment is prohibitively expensive.
Scaling generalist robotic policies is limited by the high cost of collecting embodied interaction data. Large-scale egocentric human videos offer a rich alternative, but they lack robot action labels, and existing action annotation methods are often noisy under real-world visual complexity and diverse manipulation dynamics.
A preprint from Chinese humanoid maker Galbot and the Chinese Academy of Sciences tests which off-the-shelf image encoders make video-based robot brains more reliable, and finds that one strong general-purpose encoder delivers most of the gain. Their resulting recipe tops the RoboCasa humanoid tabletop benchmark at 84.7% and adds 10 points of success on three real Galbot robot tasks.
SimpleTouch is an arXiv preprint that adds a tactile expert to the π0.5 vision-language-action model, using a frozen pretrained T3 tactile encoder and training only on 50 task demonstrations with future-touch prediction as auxiliary supervision. It reaches 77.5% average success across six UniVTAC contact-rich tasks and 71.3% across four real-robot tasks, outperforming the pretrained tactile policy FTP-1 by 10.8 and 8.8 percentage points. This suggests strong contact-rich manipulation can be learned without costly tactile policy pretraining or explicit visuotactile alignment.
This preprint introduces FastOPD, a method for distilling large vision-language-action (VLA) and world-action-model teachers into compact few-step robot policies. It keeps 84% of a 3.3B-parameter teacher's LIBERO success rate with only two sampling steps, reduces inference latency by 78.1%, and on RoboTwin 2.0 lifts single-step SmolVLA success from 35.3% to 51.2%. The work matters because it may allow foundation-policy capabilities to run on resource-constrained, low-latency robot hardware.
This preprint introduces MobiAgent, a dual-loop framework for long-horizon mobile manipulation. It decouples high-level VLM planning from low-level flow-matching skills and adds an automated lifelong learning loop from deployment data. The abstract reports no quantitative results, but the approach matters because it targets compounding errors and interference between locomotion and arm control.
Preprint. The authors build an inference-time defense for vision-language-action (VLA) policies against adversarial image patches by training a sparse autoencoder, identifying an internal feature that tracks patch presence, and suppressing that feature only when a linear probe detects an attack. On LIBERO-10 under intermittent UADA attacks, conditional suppression raises success rate, e.g. from 26.0% to 32.4% for π0.5 at n=10, while avoiding the large benign-condition drop caused by continuous suppression. This points to lightweight, fine-tuning-free robustness layers for deployed robot policies.
This arXiv preprint describes ChunkVLA-AM, a deployment pipeline that adapts an OpenVLA-OFT vision-language-action model to a FAIRINO FR3 robot for additive-manufacturing post-print retrieval. The robot predicts and executes 8-step chunks of 7-D Cartesian actions in a cloud-edge loop and completed 39 of 42 physical pick-and-transfer trials, a 92.9% success rate. The work provides a concrete feasibility check for using VLA-based automation in a fixed manufacturing workcell.
UniWAM is a preprint introducing a unified robot foundation model that combines a vision-language reasoner, a video-generation world model, and an action predictor. It reaches 99.2% success on LIBERO, 92.6% on LIBERO-Plus OOD perturbations, 68.32% on RoboTwin 2.0 Clean2Random, and 82.5% real-robot instruction-following rate. The work matters because it shows how co-training on robot, human egocentric, and VQA data can improve robustness, generalization, and long-horizon manipulation.
DynaHarness is an arXiv preprint that adds a dynamic physical harness around a frozen vision-language-action policy: a slow Qwen3-VL-4B planner proposes capabilities, while a fast 2 Hz governor grounds, monitors, substitutes and recovers them. On LIBERO-Pro it reaches 75.2% success on 800 newly sampled states versus 17.5% for the frozen π0.5 policy, and its self-evolution gains transfer to new states. This gives robot-agent builders a retraining-free way to govern and improve existing policies from execution evidence.
This preprint presents FailBank, a framework that turns an observe-only safety teacher's runtime corrections into training records for vision-language-action policies. On VLA-Arena static-obstacle tasks, it raises task success by 8.5 and 6.9 percentage points over base policies for two VLA backbones while cutting policy-induced cumulative cost by 35.6% and 23.8%. The method matters because it lets robot policies learn a persistent balance between completing a task and avoiding unintended contact, instead of relying on temporary runtime shields.
This preprint identifies a failure mode in vision-language-action (VLA) policies where instructions retrieve familiar trajectory families but do not bind language to scene-dependent action choices, so robots can execute a clean but wrong task. It introduces Equivariant Counterfactual Training (ECT), a paired data and loss scheme that supplies same-instruction demonstrations with different scenes and trains counterparts jointly. In controlled LIBERO-PRO tests, mean position-swap success rises from 36% to 59%, and on a real UR5e, unseen-position success rises from 3/40 to 35/40, showing the approach reduces counterfactual task errors.
This preprint introduces GroundingPI, a 4B grounding foundation model that predicts object points and boxes as quantized token coordinates rather than using a general-purpose vision-language backbone. Across 34 grounding benchmarks it averages 73.68%, ahead of a larger GPT-6 Astra baseline at 71.54%, and as a visual backbone it improves downstream manipulation (RoboTwin 2.0, RoboCasa-GR1) and autonomous driving (nuScenes L2 0.296 m). The result matters because it argues grounding is a distinct perceptual layer that can make embodied foundation models more precise.
This preprint introduces a framework for scene-scale aerial manipulation on an articulated uncrewed aerial manipulator, combining synthetic whole-body VLA training, measured-progress-aligned trajectory realization, and scene-graph-guided mission composition. In simulation, local skills achieve 39/60 successes under oracle handoff, and full multi-site missions achieve 21/50 (42.0%), while a monolithic whole-task VLA baseline achieves 0/50. Physical trials validate representative tasks without platform demonstrations, reducing risk and data cost for aerial manipulation.
This preprint introduces PrefPI, an iterative framework that steers pretrained generative robot policies toward user-preferred behaviors using only relative preference labels on the robot's own rollouts. On a real-robot cube pick-and-place task, it lifted object transport height from 10.7 cm to 19.8 cm after 150 labeled trajectories and cut completion time from 19.8 s to 10.8 s. It matters because it shows preference feedback can push policies outside their original behavior distribution without new demonstrations or reward functions.
This preprint reports a two-stage non-uniform Flow Matching denoiser that cuts the π0.5 VLA's action-generation steps from 10 to 2 and model-inference time from 61.6 ms to 22.0 ms, a 2.8× speedup. The authors also built a distributed real-time VLA execution framework and benchmarked six action-scheduling methods on a 180-trial bimanual physical garment-folding task; Legato achieved 96.7% success as the best training-based method, while Temporal Smoothing led training-free methods at 76.7%. Combining the fast sampler with those methods gave 1.68–2.80× inference speedups at a 6.7–10.0 percentage-point success loss, highlighting a practical speed–quality tradeoff for real-time robot policies.
This preprint introduces EWAM, a unified embodied model whose action tokens asymmetrically attend to vision-language, video, and action experts. It reports 72.1% clean-to-random success on RoboTwin 2.0, a 92.9% in-domain average, 98.8% on LIBERO, and top real-robot results across three embodiments. The work matters because it shows a learned depth-wise handoff from semantic understanding to visual foresight to action can improve robustness and transfer, including from low-cost human egocentric data.
Vision-language-action (VLA) models have advanced robotic manipulation, but their zero-shot generalization in new tasks and environments remains limited, and their reliance on specialized training keeps them from benefiting directly from rapidly advancing general-purpose vision-language models (VLMs). In parallel, recent agentic robotic systems leverage VLMs for high-level reasoning or coding agents for robot control, but often depend on extensive external models and tools, introducing additional complexity and cost.
Papers come from arXiv robotics feeds. Where our summary is not written yet, you see the opening lines of the abstract.