Technical Directions
Robotics coverage relating to open artifacts.
422 selected reportsLast 30 days 417 recordsAll coverage 1,047 records
VicOne LAB R7 explores how adversarial inputs can alter robot behavior, challenging safety assumptions. The article discusses risks from manipulated sensor data, cyber attacks, and the importance of testing safety systems against deliberate threats. It also outlines five key questions for safety and security teams to ensure robot behavior remains safe under adversarial conditions.
This research presents RynnVLA-Latent, a framework that converts heterogeneous interaction videos into scalable supervision for robot policy pretraining. By learning continuous, 3D-grounded latent actions from visual transitions without requiring native robot controls, the system achieves high success rates on benchmarks like LIBERO (98.8%) and LIBERO-Plus (84.7%), as well as strong performance on real-world manipulation tasks across dexterous-hand and mobile-gripper platforms.
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.
Zhiyuan Robotics has deployed the G2 Max, an industrial-grade embodied robot, in JD Logistics' Zhi Wang warehouse for inbound handling tasks. The robot can carry up to 50kg with high precision, achieving over 99% operational success rate. This marks a real-world application of embodied intelligence, with JD Logistics emphasizing the importance of industrial scenarios for testing and refining such technologies.
This research addresses nonsmooth action oscillations in deep reinforcement learning policies that hinder physical robot deployment. The authors challenge the assumption that temporal regularization cannot provide spatial smoothness under observation noise, proving it bounds expected action differences between states sharing a next state. They introduce CATS (Conditioning for Action using only Temporal Smoothness), which integrates a temporal penalty with linear ramp-up to learn rewarding behavior before progressively smoothing actions. Experiments demonstrate that CATS substantially reduces action oscillation without degrading task performance and with minimal computational overhead.
Neura Robotics, a German humanoid robot company, is shifting focus from hardware production to training infrastructure. Their TUM RoboGym and Neuraverse platform aim to systematize the teaching of physical tasks, enabling scalable and reusable skills across different robots and industries. The company emphasizes the importance of real-world data and physical interaction in training robots for industrial applications.
110 humanoid robots completed complex synchronized actions in 5 minutes and 43 seconds, successfully breaking the Guinness World Record. The system achieved high levels of autonomous operation, including autonomous positioning, dynamic formation, and task scheduling, marking an important step toward the practical application of humanoid robots.
This paper introduces OpenTac, a novel resistive tactile skin and full-stack ecosystem designed for customizable, scalable, and transferable 3D tactile sensing.
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%.
ZeroDex is a zero-shot framework for long-horizon dexterous manipulation that grounds language instructions into executable 3D task plans using calibrated multi-view RGB images. It extends vision-language model (VLM) grounding from 2D to 3D by fusing multi-view data, enabling contact points, placement targets, and tool-motion anchors. The system executes complex instructions by decomposing them into atomic tasks using a small library of primitives, and outperforms zero-shot visual-prompting and constraint-based baselines. It also demonstrates robustness to intermediate failures and generalization to novel objects and tool-use tasks.
This paper introduces IMPULSE, a technique that enables latent world models like V-JEPA 2 to infer object physical properties through brief calibration pushes. By using a Bayesian contact memory and a lightweight random-Fourier pose head, the method reduces 1-second feature-forecast error by approximately 11.6–12.8% on fresh objects in held-out evaluations across multiple training seeds. The approach demonstrates that motion-based pathways are necessary for physics to reach features, with gains scaling after contact onset and extending to out-of-range physics scenarios.
This research proposes Collective Experience Learning (CEL), a paradigm where deployed robots pool execution experience to update a shared policy. The method, implemented via Failure-Guided Experience Retrieval Fine-Tuning (FERT), uses failure contexts to retrieve successful trajectory segments for supervised updates without requiring new expert demonstrations. Evaluated across five real-world manipulation tasks on robots excluded from training, the approach achieves success rates between 93.3% and 99.2%. Specifically, on multi-stage phone packaging, using a fleet of 96 robots improved success rates from 79.2% to 98.3%, validating CEL as a scalable method for improving robot policy reliability through deployment-generated experience.
At the 2026 China International Import Expo, Faoro Robotics showcased its industrial-grade force-controlled humanoid arms (ART series), embodied AI robots (BW10-Lite), and smart workstation solutions for tasks like shoe manufacturing and box folding. The company announced it has secured nearly 1 billion yuan in C+ and D round funding from investors including Meituan Dragon Ball, along with five strategic partnerships to advance its industrial autonomy ecosystem.
This research introduces the MAP (Masked Action Space) interface to unify spatiotemporal control references for robot manipulation. The proposed Inpainting Policy (IP) reconstructs missing action dimensions using progress-guided alignment and multi-mask training, achieving an average task success rate of 92.67% on ManiSkill and 64.67% on LIBERO-10. Real-world experiments show average positional and rotational errors of 44.3 mm and 7 degrees, respectively, validating the policy's ability to guide executable robot behaviors using diverse references.
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.
This research paper introduces a 'wind memory' mechanism to improve odour source search strategies when wind direction reverses.
This research presents a mode-unified control framework for robotic lower-limb prostheses that represents knee and ankle biomechanics continuously as functions of terrain slope and gait phase.
This paper presents SPARK, a framework that maintains manipulation plans by anchoring symbolic behavior trees to current 3D observations, enabling local recovery from failures when objects move. Evaluated on LIBERO-PRO without ground-truth poses or task-specific practice, the system achieves 75.1% mean success across position and task settings and 69.7% on held-out language and object settings. It also demonstrates 68% success across 11 physical task-embodiment settings on three different robots using an open-weight 27B planner with 4-bit weights.
At the 2026 China International Import Expo, Realtek Semiconductor highlighted its communication and control components for humanoid robots, including EtherCAT MCUs, single-pair Ethernet PHYs, TSN-enabled industrial switches, audio DSPs for noise cancellation, and synchronized wireless camera solutions. The display emphasized that reliable data transmission, precise timing, and robust industrial networking are essential for coordinating robot joints, processing sensory inputs, and ensuring stability in factory environments. Realtek's approach extends its established networking technologies to new robotic terminals, noting that successful deployment depends on system integration and long-term validation rather than component specifications alone.
This paper presents OmniDex, a benchmark and model for dexterous hand grasping in cluttered scenes. It addresses data scarcity through a scalable seed-and-filter strategy, generating over 2.6 million scenes and 0.4B grasp ground truths. The model combines Soft Winner-Takes-All learning with physical constraints to achieve robust grasping without post-optimization latency, showing strong generalization across diverse scenarios.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).