Skip to content

Technical Directions

VLA Latest coverage

Robotics coverage relating to vla.

367 selected reportsLast 30 days 343 recordsAll coverage 758 records

updated

Signal archive — Page 17

2026-10-27Tue321–340 records
  1. arXiv Robotics — research abstracts75

    PS-DLO: Property- and Spatially Aware Deformable Linear Object Manipulation Policy

    This paper introduces PS-DLO, a manipulation policy that combines contrastive image-text learning to infer object properties with SAM3-based mask-guided feature extraction for future observations. Evaluated across nine real-world tasks on single-arm and bimanual platforms, the shared multi-task policy achieves an average success rate of 68.9%, surpassing the strongest baseline by 20.0 percentage points. In generalization experiments, PS-DLO outperforms the baseline in six of seven visual-shift conditions for butterfly-knot untying, demonstrating improved robustness to visual variation when combining object-property priors with future spatial supervision.

  2. arXiv Robotics — research abstracts75

    MOON: Learning When to Forget and How to Move On in Vision-and-Language Navigation

    This research presents MOON, a framework that equips frozen vision-and-language navigation policies with lightweight modules for detecting deviations and recovering from them. By clearing visual history upon deviation detection and training a recovery policy via GRPO, MOON improves success rates on R2R-CE and RxR-CE benchmarks by 5.2 and 6.0 points respectively, outperforming compared methods without retraining the base policy.

2026-10-26Mon
  1. arXiv Robotics — research abstracts75

    SAGE: A hybrid vision-language and geometry framework for work-zone pathway prediction under temporary traffic control

    This paper introduces SAGE (Semantic-to-Action Grounded Execution), a read-decide-execute-check framework that integrates vision-language model trajectories with geometric path candidates to navigate temporary traffic control zones. Trained on 2,306 real-world ROADWork validation episodes using MUTCD-compliant supervision, SAGE reduces average displacement error by 8.3% and final displacement error by 12.0% relative to a geometric ranker, while reducing hard-constraint violations by 37%. The system achieves a path within 10 pixels of human demonstrations in 87% of episodes and outperforms four other adapted driving VLMs in safety measures.

  2. arXiv Robotics — research abstracts

    Full-Weight VLA Fine-Tuning Without an fp32 Master Copy Silently Erases Most Updates

    This paper reveals that fine-tuning VLA models with bfloat16 weights without an fp32 master copy leads to significant loss of updates due to rounding. The study shows that retaining fp32 weights during training improves performance on tasks like LIBERO-Object, with statistically significant results across multiple models and training configurations.

2026-10-23Fri
  1. arXiv Robotics — research abstracts82

    Whole-Body Motion Guided Reference Generation for Uncertainty-Aware Compliant Assistance

    This research proposes a dual-loop assistance strategy for exoskeletons that couples multimodal trajectory generation with variable admittance control. The method uses whole-body human motion and exos, and incorporates task descriptions as semantic cues. A three-stage training pipeline reduces cross-modal representation mismatch, while trajectory variance provides an uncertainty measure for prediction reliability. Experimental results show reduced muscle activation and metabolic energy expenditure compared to standard baselines.

  2. arXiv Robotics — research abstracts75

    Memory before Goal: Role of Pre-History Memory for Humanoid Loco-Manipulation

    This paper presents WholeBodyMem, an object-guided perceptual memory system that retains pre-history information for humanoid loco-manipulation in dynamic environments. By using SAM3 to track task objects and causally selecting key frames within an adaptive budget, the action expert integrates these frames via cross-attention with object tokens encoding appearance, geometry, and time. The authors also release HuMeGo, a benchmark featuring six simulated and six real-world Unitree G1 tasks. WholeBodyMem improves success rates from 29.1% to 72.6% in simulation and from 36.7% to 88.3% on the real robot, while also enhancing performance on standard SIMPLE tasks from 70.2% to 86.9%.

  3. arXiv Robotics — research abstracts75

    Multimodal Progress-Residual Fusion for Anticipating Navigation Retreat in Robotic Flexible Endoscopy

    This research presents Multimodal Progress-Residual Fusion (MPRF), a causal model designed to anticipate operator retreat decisions during robotic flexible endoscopy. By fusing electromagnetic tracking, endoscopic video, and motor intent data against a nominal-progress baseline, the system predicts navigational mismatches caused by friction or deformation. Evaluated on a 150-episode colon-phantom dataset, MPRF achieves an AUPRC of 0.580 +/- 0.007, outperforming reduced-modality controls (0.511). The work establishes a foundation for behavioral assistance systems in operator-controlled navigation.

  4. arXiv Robotics — research abstracts75

    Self-Distilling Future Video Representations for World Action Models

    This research presents FRED, a framework that enhances World Action Models by using self-distillation to predict future video representations from noise in a single denoising step. By conditioning on historical observations and using ground-truth future representations as targets, FRED improves action prediction quality. It demonstrates significant gains in long-horizon and out-of-domain tasks under zero-shot generalization, achieving 50.0% task success on tasks absent from robot training data but demonstrated in human videos, compared to 19.4% for the best competing World Action Model.

2026-10-21Wed
  1. Universal Robots — Official announcements68

    Teradyne Robotics and Sereact Partner to Scale Physical AI Solutions for Retail and Logistics

    Teradyne Robotics and Sereact have announced a partnership to deploy AI-powered physical solutions in retail and logistics, using Universal Robots' arms with Sereact's Cortex platform. The collaboration includes live deployments at Zalando's fulfillment centers, demonstrating the ability to handle unstructured tasks like returns processing. The solution runs on NVIDIA technologies, including Isaac Sim and Jetson Thor, and is being scaled globally.

  2. arXiv Robotics — research abstracts82

    PACT-VLA: Speculative Action-Chunk Verification with Pre-Commit Observations

    This paper presents PACT-VLA, a training-free framework that improves the success rate of vision-language-action (VLA) policies by using pre-commit observations to select the best action chunk during inference delay. It outperforms existing methods on both simulated and real-robot tasks, achieving significant improvements in task success rates.

2026-10-20Tue
  1. arXiv Robotics — research abstracts82

    VLA-Roll, Fewer Denoising Steps, Better Control via Adaptive Action Reuse

    This research proposes VLA-Rolling, a framework that reuses action tails from vision-language-action models to enhance control efficiency. By adapting denoising budgets and refining actions through confidence-guided steps, the method achieves faster inference and improved success rates across multiple benchmarks.

  2. arXiv Robotics — research abstracts82

    EndoVLA-OOV: Language-Conditioned Out-of-View Target Acquisition in Robotic Endoscopy

    This research presents EndoVLA-OOV, a vision-language-action policy for robotic endoscopy that translates anatomical descriptions into camera motions. The policy uses a pretrained backbone adapted to continuous distal bending commands and combines images to guide search. An open-source dataset with 1,050 episodes and 262k observation-action pairs supports policy learning. Physical evaluations show 93.3% success in cross-embodiment transfer, with success rates improving from 40.0% to 90.0% with more training data.

2026-10-19Mon
2026-10-18Sun
  1. Chinese robotics — Robot learning research75

    From 'Understanding the World' to 'Simulating the Next Step': How World Models and VLAs Give Robots a True 'Embodied Brain'?

    This report examines the convergence of VLA models and world models in robotics, highlighting deployments in pharmacy and industrial settings. It details specific technical achievements, including a model achieving a 2.6x improvement in success rates on unfamiliar scenes with 30,000 hours of training data, and discusses industry standards being developed to evaluate these technologies.

2026-10-17Sat
  1. QbitAI — Robotics72

    From Understanding the World to Predicting the Next Step: How World Models and VLA Enable True Embodied Intelligence in Robots

    The article discusses the development and deployment of world models and VLA (Vision-Language-Action) in robotics, focusing on their roles in enabling autonomous decision-making and physical interaction. It highlights real-world applications in retail and manufacturing, along with ongoing research and standardization efforts in Shanghai.

2026-10-16Fri
  1. Chinese robotics — Robot learning research82

    GPT-6-Astra's Embodied Intelligence Put to the Test, Traditional Benchmarks Score Perfectly, Real Robots Hit Snags

    This research paper evaluates GPT-6-Astra's embodied intelligence capabilities across 12 simulated benchmarks and 8 real-world robot platforms. While it achieves high scores in structured tasks like LIBERO-Long++ and SIMPLE, real-world execution reveals significant limitations, including safety issues, mechanical load constraints, and failure to handle complex, unstructured tasks. The study highlights the gap between simulation and physical deployment, and proposes improvements through better execution systems and task-specific adaptations.

  2. arXiv Robotics — research abstracts75

    X-Fly: Learning to Reach and Stop in Full-Chunk Aerial Vision-Language Navigation

    This research presents X-Fly, a staged training framework for full-chunk aerial vision-language navigation. By executing complete predicted action chunks before replanning, the method reduces model inference time from 195 ms to 42 ms per waypoint and improves recovery from non-expert states. The approach combines Recovery DAgger labeling with Flow-GRPO refinement to jointly optimize a chunked flow controller and a Bernoulli STOP policy. On the OpenFly test-seen dataset, X-Fly achieves a 74.18% success rate compared to 16.45% for full-chunk behavior cloning and 32.04% for DreamFly.

  3. arXiv Robotics — research abstracts75

    A Phenomenology of Neural Geometry in Vision-Language-Action Models

    This research paper investigates the internal geometric structure of vision-language-action (VLA) policies used for robotic manipulation. By analyzing activation patterns through controlled sweeps and linear probes, the authors establish a taxonomy of seven task-relevant concepts. The findings reveal that these concepts are predominantly realized as multi-dimensional shapes influenced by physical constraints, such as wrist orientation forming an open arc due to joint limits rather than a full circle. The study confirms these geometric realizations across different VLA models on real-world robot data and demonstrates actionable methods for steering along these specific geometries to influence model behavior.

2026-10-15Thu
  1. arXiv Robotics — research abstracts78

    LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting World

    This research presents LEGS, a hybrid simulator that combines a mesh foreground with a photorealistic 3D Gaussian Splatting background to train vision-language-action (VLA) policies for humanoid loco-manipulation without human teleoperation. The method uses a procedural motion-primitive generator to synthesize labeled demonstrations and a two-stage color calibration to align renders with the deployment camera. Policies trained on LEGS data match or exceed those trained on human teleoperation demos and outperform mesh-only baselines. The approach enables cost-effective re-rendering of demonstrations for new scenes and objects.

  2. arXiv Robotics — research abstracts82

    Robota-Embed: Unified VLA Embeddings for Robot In-Context Learning and Evolution

    Robota-Embed is the first vision-language-action (VLA) embedding model that jointly encodes task, execution stage, and future continuation for robots. It achieves 47.33% clip-level and 81.77% episode-level R@1, outperforming action-free Gemini Embedding 2. The model supports training-free and training-based policy conditioning, improving success rates in simulated and real-world tasks, and enables task-local context evolution with reduced policy token usage.

Timezone · UTC

Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).