The authors propose Zero-real Agent Policy (ZAP), which uses a pretrained 27B-parameter vision-language model with a simple tool interface to control a robot, trained via supervised fine-tuning and GRPO on simple simulated tasks, without using any real robot data during training.
This work proposes a 3D Gaussian Splatting (3DGS)-based data generation pipeline for adapting wrist-camera VLA policies to fine reaching tasks. The method reconstructs the real workspace and renders wrist-camera observations from robot-aligned poses, pairing them with smooth, structured reaching trajectories for VLA fine-tuning, and adds an instruction-conditioned reach verifier that judges termination before physical contact.
RoboDoctor is a training-free framework that, under fixed model weights, lets generalist robot policies and language-model agents adapt during execution through three revision targets: diagnosis, action execution, and persistent experience records. On 54 RoboDojo-Sim task entries, the framework achieved a five-dimension mean score of 57.10 and a native success rate of 49.44%; recorded interactions show policy switching, selective correction, and grasp revision. The authors note that these results characterize execution-time adaptation in the evaluated system and that controlled experiments are still needed to isolate the contribution of cross-episode experience.
This study proposes a task decomposition framework for deep reinforcement learning (DRL) that combines transfer learning (TL) with reward redesign (RR) to improve learning efficiency and robustness in obstacle environments.
The authors propose DEX-PHD, a portable handheld device that integrates a dexterous robot hand, continuous per-finger control, and multimodal logging, and can collect proprioceptive information, fingertip contact data, and wrist-camera video, with the operator directly controlling the robot hand through joint-level mapping. Its untethered design supports data collection in a variety of environments. The authors demonstrate it on contact-rich tasks that require multi-finger coordination, and execute the collected demonstrations on a physical humanoid robot through open-loop replay; the results show that open-loop replay preserves the main temporal patterns of finger motion and task-relevant contact.
The paper proposes that semantic failure analysis is a key step toward mature dexterous manipulation of deformable objects and supports this claim with proof-of-concept experiments. Unlike rigid-object manipulation, cables, plastic parts, cloth, and similar objects exhibit diverse and hard-to-detect failure modes such as buckling, plastic deformation, pressure changes, and tearing due to differences in materials, geometry, and behavior, requiring different corrective strategies. The authors define the components and challenges of semantic failure analysis and present experimental results indicating its potential as a research direction.
Zero-HLM proposes a real-to-sim-to-real framework that turns smartphone captures and simulated source demonstrations into scalable training data for humanoid loco-manipulation. The study generated 4,800 physically validated demonstrations and achieved zero-shot transfer on three Unitree G1 tasks; GR00T N1.7 trained in simulation reached a mean full-task success rate of 66.7%, versus 46.7% with real-data training.
This study proposes using factor graphs to probabilistically model deformable linear objects (DLOs), estimating the magnitude, direction, and location of 3D forces acting on them while simultaneously reconstructing the cable's shape. Because cables are thin and unactuated, strain gauges and actuation readings are generally impractical; the method uses only one type of sensing information and performs force estimation in simulation tests that add measurement noise and have the cable move along a prescribed path. The authors state that this general method can apply to a broader range of tasks.
Researchers propose JevVLA, a lightweight energy-based adaptation framework for vision-language-action (VLA) policies that, with the base policy frozen, learns a differentiable scalar energy function used both for comparing discrete candidates and for correcting or generating continuous actions.
MMGNet proposes a multimodal grasp synthesis network that uses hierarchical image-to-3D feature injection and rotational graspness modeling to integrate multi-scale image features into 3D representations and jointly predict grasp approach direction and in-plane rotation. On GraspNet-1Billion with RealSense data, it achieves a mean AP of 60.34, surpassing all compared methods; in real-world experiments with adversarial grasp objects, its grasp success rate is 8.46 percentage points higher than AnyGrasp. Ablation studies show that the two modules provide complementary gains. Code, models, and data will be released after the paper is published.
The authors propose PhysLadder, a hierarchical sim-real benchmark for evaluating physical interaction capabilities in contact-rich manipulation, comprising 16 household tasks divided into three cumulative levels: contact selection and geometric alignment (L1), force regulation under sustained contact (L2), and coordinated contact reconfiguration and force regulation (L3).
ARPG proposes separating action representation learning from action generation: first pretrain an action representation encoder on large-scale human manipulation data, then freeze that encoder and train a 130M-parameter generative backbone in its representation space to learn task-conditioned action distributions.
AuRA is an automated robot interaction and data preparation system for digital-twin reconstruction, combining execution-aware grasp selection, gripper-motion measurements derived from robot states, and registered pre-interaction geometry and appearance priors. Across 13 deformable objects, it reduces active human labor and wall-clock collection time per successfully reconstructed asset by 88.5% and 55.4%, respectively; compared with the human-collected PhysTwin baseline, it improves 7 of 9 metrics for future prediction under known gripper motion, including a 28.1% reduction in one-way L1 Chamfer distance.
The authors propose ThinkLocoManip, a whole-body control agent framework for legged loco-manipulation that uses an interleaved reasoning-and-acting paradigm to handle long-horizon tasks in large-scale environments. The framework orchestrates perception, navigation, manipulation, and locomotion as composable tools, and it designs cross-domain coordination skills and a keyframe thinking mechanism to support error recovery. In real-robot experiments, its success rate was 77%, more than double the 37.1% of the strongest plan-then-act baseline.
The research team proposes the RynnVLA-Latent framework, which uses RynnLAM to learn continuous, 3D-grounded latent actions from visual changes without native robot control signals, converting 30.1K hours of human, robot, and simulation video into 745.52M latent action descriptors for policy pretraining.
Why it matters: This research uses 3D-grounded latent actions to turn heterogeneous interaction videos into pretraining supervision, offering a reusable path for VLA models to bypass the scarcity of action annotations.
AdaPolicy proposes an action-driven anticipatory policy framework that shapes future-aware representations through cascaded next-chunk prediction, requiring no encoding of future observations during training and adding zero inference-time overhead. The method achieves consistent gains on MetaWorld, RoboTwin 2.0, LIBERO, and real-world deployment, with improvements of up to 14 percentage points on millimeter-precision tasks such as headphone assembly and shoelace threading.
HuGe is a family of humanoid generalist vision-language-action (VLA) models pretrained on over 2,000 hours of egocentric human and robot data. It predicts latent motion tokens to drive a pretrained whole-body controller and outputs hand-control signals.
The authors propose PS-DLO, a deformable linear object (DLO) manipulation policy that combines object properties with spatial awareness. Through contrastive image-text learning, it maps initial observations to categorical descriptions of DLO properties, and uses SAM3 to extract mask-guided features from future observations, which are distilled into an action expert. On nine real-world tasks across single-arm and bimanual platforms, PS-DLO achieves an average success rate of 68.9%, 20.0 percentage points higher than the strongest baseline; in six of seven visual-shift conditions for butterfly-knot untying, its success rate exceeds the baseline.
ElevateHead proposes an action head that selects the task execution mode once at a learned decision point and keeps it fixed for the entire episode; users can also set this choice as an input. On held-out Franka Kitchen tasks, 45.2% of its episodes completed all four subtasks, versus 29.3% for BeT (+0.160, 95% CI [+0.129, +0.188]); on held-out block stacking tasks, version 4 was comparable to BeT but did not surpass it.
The authors propose a task and motion planning (TAMP) framework that alternates task-directed interactive information gathering with planning and execution. It plans optimistically for interactive actions based on PDDLStream, verifies outcomes during execution, retains actions that remain applicable, and replans only the subsequent connections. The framework was tested in simulation and on a real robot on tasks such as determining whether an object is movable, whether a pen has ink, whether a drawer can be opened, and whether a container is empty. The results show that interaction-driven state updates and recovery replanning can support continued execution under uncertain outcomes and external disturbances.
RoboLongR proposes an offline-to-online reward learning framework that uses causal consistency to align reward progress with the prerequisites supported by execution history, guiding policies from behavior cloning to online reinforcement learning. Pretrained on 88 RLBench tasks and Open-X data, the method achieves an average success rate of 31.8% on 18 long-horizon tasks excluded from reward pretraining, 9.2 percentage points above the strongest comparison baseline; on four real-robot tasks it achieves a 57.5% success rate, 32.5 percentage points above the comparison method.
AdaDexTrack proposes a modulator-in-the-loop framework for language-conditioned dexterous manipulation tracking, using three types of feedback correction—reference modulation, object-latent modulation, and positional-target modulation—to mitigate synthetic reference noise, embodiment bias from human-to-robot retargeting, and error accumulation in fixed-reference tracking.
The authors propose the Open-World Law-to-Action (L2A) framework, in which a vision-language model proposes, for each movable part, its motion type, the terms of its law of motion, and possible locking factors; the part's law is then identified online via Bayesian linear regression with closed-form evidence, and touch can overturn the priors.
A survey reviews 113 learning-based dexterous manipulation methods, covering the dexterous hands used, the training simulators, and how results are reported. The authors note that nearly half of the works use the same recipe: reinforcement learning on privileged state in a GPU simulator, then distillation into a vision-only student model transferred to hardware, and nearly all of them concentrate on the same simulator.
The authors propose Multi-constraint Dexterous Policy Optimization (MDPO), which formulates dexterous manipulation as a multi-constraint policy optimization problem involving both cumulative cost constraints and state-wise cost constraints, and integrates multi-constraint projection into the policy optimization process. In experiments on ReDMan for safe joint control, finger safety, and grasping tasks, MDPO achieved better constraint satisfaction than baselines while maintaining competitive task performance. The code is publicly available.
MOTIO is an agentic skill engineering system that turns ordinary human action videos, realistic 3D assets, motion references, and reusable robot skills into executable humanoid loco-manipulation programs, without task-specific teleoperation or robot demonstration collection.
The authors propose Tiny Recursive Policy (TRP), a parameter-efficient behavior cloning policy that repeatedly updates high-level and low-level latent states using a shared computation module at each control step before decoding an action sequence, and propagates the state to subsequent predictions.
The authors propose ORBIT (Orbital Reconfigurable Base-actuated Integrated Theta-motion Hand), a modular multi-finger robotic hand platform whose core is a rotating knuckle mechanism: fingers travel around a cylindrical palm via a belt drive, enabling finger-layout reconfiguration within 1 second without disassembly.
In 2025, Joyson Electronics adjusted its business direction to a "automotive + robotics" dual-drive strategy, positioning itself as a Tier 1 supplier in the robotics field. Its product plan covers key components such as AI head assemblies, energy management systems, brain and cerebellum domain controllers, electronic skin, and dexterous hands.
Geluo showcased four semiconductor-specific AMRs with complementary positioning, aimed at existing 200mm (8-inch) fabs, enabling incremental automation adoption without large-scale infrastructure modifications.
Mech-Mind will exhibit seven "Eye-Brain-Hand" intelligent applications at VIMF 2026, held November 4-6 at the Kinh Bac Cultural Center in Bac Ninh Province, Vietnam. These cover high-precision assembly of FPC flexible flat cables, flexible picking and precision assembly of graphics card power cables, solder joint defect inspection for mobile phone flash units, appearance defect inspection for Type-C connectors, Pin measurement for automotive electronic connectors, screw height inspection for laptop cooling plates, and machine loading/unloading of high-reflectivity gears, targeting precision manufacturing scenarios such as 3C electronics, semiconductors, automotive electronics, and machining. Booth: Hall E, 350.
Shenmou Intelligent's "Dragon Claw" aerial-operations drone has won State Grid Power's "High-Stability UAV for Aerial Operations" project, using a drone with a robotic arm and dexterous hand to clear foreign-object hazards on overhead transmission lines. The system identifies foreign objects such as bird nests through inspection, then uses the robotic arm and dexterous hand to handle and verify them, integrating detection, handling, and verification into one workflow. Shenmou says its core technologies include a self-developed flight control system, robotic arm and dexterous-hand grasping control, dynamic visual guidance, and piezoelectric six-axis force sensors, and cites research estimating annual procurement demand of about 150 million to 200 million yuan per province.
The study proposes Quality-Controllable Policy (QCP), which enables inference-time control over individual quality attributes such as execution speed and motion smoothness, as well as their combinations. To extend the controllable quality range beyond the initial demonstrations, the authors designed an iterative hindsight relabeling loop: collect online rollouts under target quality labels, add successful trajectories to the dataset, then relabel the accumulated trajectories in hindsight according to their measured quality and fine-tune the policy.
Afferent-Omni is an open-source visuo-tactile command and control (C2) framework for embodied AI and dexterous manipulation, offering a FastAPI backend, 9 modular routers, 15 REST endpoints, and a 50 Hz closed-loop WebSocket stream.
German supermarket group Edeka has added two Knapp Pick-it-Easy robots to its Nordfrische meat-processing distribution center in Lüttow-Valluhn for fully automated picking of fresh meat products.
Agibot released the safety-certified heavy-load embodied robot Genie G2 Max at its AI WAVE event on October 26. It features 7-DOF force-controlled dual arms, with a single-arm payload of 20 kg, a dual-arm rated payload of 40 kg, a peak payload of 50 kg, and repeat positioning accuracy of ≤±0.5 mm.
Zhongke Huiling has completed an A++ funding round worth several hundred million yuan, jointly invested by Zhilai Investment, Yonghua Investment, Maotian Investment, Yuanhe Chenkun, Guoke Kezhi, Luning Investment and others. The funds will be used for product mass production and delivery, R&D and iteration of embodied brain technology for key scenarios, and market expansion.
Zhang Changshui and Cui Sen's team at Tsinghua University propose RoboCoach, an embodied intelligence system that uses the world model CoachWorld to run closed-loop trials of long-horizon tasks in imagination, identify the first weak subtask that times out, and request targeted human teaching for it, updating only the corresponding skill expert. Under a fixed demonstration budget, the success rate rises from 13.3% to 75.0% on Franka and from 40.0% to 83.8% on AgileX, while the uniform collection baseline reaches only 30.0% and 47.5%, respectively. On four unseen new task routes, the average success rate is 35.0%, compared with 0% for the shared-policy baseline.
Why it matters: The paper presents a complete experiment on using a world model to locate weak subtasks and improve success rates through targeted teaching; its approach to data budget allocation may be a useful reference.
The Samsung RX Lab-China team proposes ε4P (Imperfection for Precision), which admits imperfect data into different noise intervals of flow matching according to the type of data defect, thereby turning data that is harmful when mixed directly into training into a performance gain.
At the 26th China International Industry Fair, Inovance unveiled its wheeled humanoid robot InoBot W1 for the first time and connected it to an automotive welding demonstration production cell, where it works with a vision system and an arc welding robot to autonomously complete tasks such as workpiece loading and unloading and resetting.