Technical Directions
Robotics coverage relating to vla.
286 selected reportsLast 30 days 265 recordsAll coverage 622 records
This paper presents TouchBridge, a framework for transferring human demonstrations to robot manipulation using tactile sensing and contact alignment. The method enables direct conversion of human demonstrations into robot-domain training data without requiring object meshes or physics simulations, achieving 72.5% success rate in real-world tasks.
This paper presents RoboSeed, a framework for evaluating vision-language-action (VLA) and world-action models across five dimensions: capability, downstream training affordability, deployment affordability, locally verified research value, and reproduction evidence. The authors apply the framework to representative models to reveal budget-dependent backbone choices and sensitivity to adaptation recipes, while noting that incomplete serving traces prevent overall numerical rankings.
This research presents PRIME W0, a hierarchical conditional flow-matching framework designed to enhance Vision-Language-Action (VLA) models by integrating long-horizon temporal context and hierarchical VLM representations. The system employs three key designs: per-layer KV injection from a pretrained VLM into an Adaptive MoE action expert, dynamics-aware conditioning via asynchronous world-model subgoal-image prediction and a DINOv2 historical-frame semantic bridge, and Action Dynamic Manifold (ADM) flow matching with covariance-structured noise that respects robot kinematic coupling. Trained in four stages including VLM pre-training, DAT/NTP visuomotor pre-training, DiT flow-matching decoder assembly with WM injection, and end-to-end SFT, PRIME W0 achieves state-of-the-art results on LIBERO and real-robot benchmarks, with systematic ablations validating each architectural component.
This paper argues that predicting robot actions is fundamentally a geometry problem rather than a language generation task. The authors show that compact Vision-Language Models (VLMs) with 0.5-2B parameters already encode rich spatial information in their intermediate hidden states. By attaching a lightweight, geometry-compatible prediction head to a frozen backbone, the method regresses affordance boxes and trajectory waypoints directly as continuous values, bypassing the autoregressive decoding bottleneck. Evaluated on ShareRobot-Bench and AGD20K, the 2B variant reduces trajectory error by over 60% and improves affordance mAP by 85% relative to an autoregressive baseline, approaching RoboBrain performance at significantly higher throughput.
This paper introduces SONIC-K1, an open-weights whole-body motion tracker adapted for the Booster K1 humanoid robot. Trained from scratch on a single shared GPU with significantly reduced compute compared to the original SONIC framework, the policy demonstrates strong performance in MuJoCo simulation and improves upon an initial real-robot release that exhibited leg shaking. The code, policy weights, and evaluation tools are publicly available.
This paper introduces OpenFlyV2, a unified onboard architecture combining vision, language, and action capabilities for aerial robots. The system utilizes a shared multimodal representation to handle semantic understanding, spatial reasoning, and continuous flight control on resource-constrained hardware like the NVIDIA Jetson Orin NX. A simulation-based dataset containing approximately 70,000 navigation trajectories and 60,000 instruction-response pairs supports zero-shot transfer to physical quadrotors, achieving a 48% success rate in four real-world environments without additional real-world aerial training.
This research investigates how an 8B-scale navigation system should organize information for zero-shot vision-and-language navigation.
This research proposes Self-Labeled Preference Transfer (SPT), a method that allows robot policies to adapt to new tasks without requiring manual preference annotations. SPT uses a shared vision-language backbone to recognize preferences from annotated source demonstrations and self-labels target demonstrations for fine-tuning. In simulation, SPT achieves 84.8% preference following and 89.2% task success, compared to 59.4% and 88.0% for naive fine-tuning. In real-world experiments, SPT improves preference following from 52.5% to 92.5% relative to naive fine-tuning across two task transfers.
At the October 29, 2026 A.I. Day forum in Shanghai, Zhiyuan presented its Physical AI technology loop, featuring the GO-3 Preview foundation model, Genie Master agent framework, and G2 Max heavy-duty industrial robot. The company highlighted engineering standards for deployability, replicability, and sustainability, citing zero-shot success rate improvements via data scaling, collective experience learning strategies that outperform remote operators in specific tasks, and real-world deployments in manufacturing lines such as Longqi Factory for loading/unloading operations.
At the 2026 Zhiyuan Innovation Day forum in Shanghai, Zhiyuan presented its transition from stage demonstrations to industrial deployment. The company unveiled the GO-3 Preview foundation model, Genie Master agent framework, and the G2 Max heavy-load industrial robot. Specific metrics highlighted improvements in task success rates through collective experience learning and post-training scaling. Early deployments are occurring in electronics manufacturing and retail logistics settings.
On October 30, the Collective Experience Learning (CEL) framework for embodied intelligence was released in Shanghai. Developed by Agibot in partnership with Tsinghua University and Wuwen Xinqiong, CEL enables robots to learn from each other's real-world operation trajectories rather than relying solely on human demonstrations. In tests involving five tasks such as flower arranging and phone packaging, robot strategies achieved success rates exceeding those of remote human operators, with the highest improvement reaching 16.7 percentage points. The framework also demonstrated that increasing the number of participating robots within the same physical training time directly improves task execution success rates, breaking previous scalability limits.
On October 29 at the Shanghai Expo Center, AGIBOT executives discussed embodied AI progress during its AI DAY event. They highlighted challenges in scaling deployment, including data efficiency, operational ratios, and industrial reliability. While initial deployments at Chimelong Paradise involved over 300 robots across more than 100 locations, the project is expanding toward 500 units. Executives emphasized that true scalability depends on reducing reliance on on-site engineers, improving model robustness through post-training, and ensuring long-term stability in complex environments.
This release offers a LeRobot policy conversion of the NVIDIA Cosmos3-Edge base checkpoint. It serves as a starting point for robot policy post-training without task-specific fine-tuning. The model jointly predicts robot action chunks and future video latents from language and current camera observations, with optional current-state conditioning. Usage requires a target dataset to resolve frame rate, camera size, and action layout. The package supports both DROID post-training recipes and LIBERO default configurations, with options to preserve or reinitialize action modules.
This repository contains the LeRobot policy conversion of the NVIDIA Cosmos3-Nano base checkpoint.
AGIBOT has released GO-3 Preview, an embodied foundation model that integrates pre-training, post-training, and post-deployment learning phases. The model reportedly achieved first-place results on three major benchmarks: RoboDojo, RoboTwin, and RoboColiseum. Built on the WRAM (World Reasoning Action Model) architecture, it aims to unify task understanding, future-state prediction, and action generation. The system includes Genie Evolver for hybrid physical-virtual training and Genie Master as an embodied agent framework. AGIBOT plans to fully open-source GO-3 Preview by the end of 2026.
This research introduces PDRVLA, a Vision-Language-Action (VLA) model designed to overcome the latency of video world models by using a lightweight Physical Dynamics Module.
This research proposes R2A, a framework that enhances the robustness of Vision-Language-Action (VLA) policies by treating out-of-distribution visual shifts as a geometric motion primitive retrieval problem. R2A uses a graph neural network to dynamically localize robot execution and harmonize retrieved geometric priors with predicted actions, achieving significant improvements in simulation and real-world benchmarks without policy finetuning.
This article categorizes companies working on embodied intelligence into four types: independent brain companies, integrated hardware-software manufacturers, tech giants and large model firms, and compute and open platform enterprises. It highlights ACE ROBOTICS as a leading independent brain company with its Kairos 3.1 world model achieving top results in multiple benchmarks. The piece also discusses challenges such as data scarcity and simulation-to-reality gaps, and provides a framework for selecting the right company based on technical, open-source, and application criteria.
This paper introduces Zero-real Agent Policy (ZAP), a framework that separates semantic reasoning from motion control to enable agentic robotics. By training motion skills entirely in simulation using simple primitives and the model's own successes without frontier teachers or real-world data, ZAP achieves 96.7% success on real-world tasks, matching GPT-6 Astra performance while running three times faster and seventeen times cheaper. It also demonstrates generalization to new objects, goals, strategies, and unseen robot embodiments.
This research presents a multimodal framework that integrates speech, vision, and touch to enable contextually appropriate robot responses. Evaluated in virtual reality with 18 participants across five social scenarios, the autonomous system was rated above neutral for naturalness and social appropriateness but received significantly lower scores than a human-teleoperated baseline. Analysis identified conversational timing, clarification of ambiguous requests, and spatial coordination as key factors contributing to the performance gap, highlighting challenges in developing autonomous robots capable of natural multimodal social interaction.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).