EmbodyHOI runs a differentiable inverse kinematics (IK) solver for a specific humanoid embodiment inside the sampling loop of a frozen flow-matching hand-object interaction (HOI) trajectory generator, placing embodiment reachability constraints within the generation process rather than in post-hoc retargeting.
An arXiv paper benchmarks nine depth estimation model families on Jetson AGX Orin, covering monocular methods Depth Anything V2 and 3, and Depth Pro.
A study proposes a plug-and-play auxiliary loss called score-decorrelation to address the problem that diffusion policies with linearly composed scores tend to have convergent component predictions that are hard to interpret in multitask robot learning. The loss encourages specialization by penalizing overlap between component noise predictions, with almost no architectural changes and very low computational overhead. Experiments on multiple benchmarks show that the method consistently reduces duplicate counting between components, improves task decomposition, and produces behaviorally meaningful specialization without compromising multitask performance.
A method for autism spectrum disorder (ASD) education enables a camera-equipped Nao robot, using a vision-based large language model, to convert photos of a Shoebox Task into a Planning Domain Definition Language (PDDL) representation and generate action plans, autonomously executing the task, prompting the next step, or reporting failures through action demonstrations, voice, and gestures.
A study embeds capacitive sensing into a gecko-inspired dry adhesive interface to recognize five interaction states in real time during on-orbit servicing grasps (idle, proximity, contact, engaged, disaged). Random Forest achieved 95.65% offline classification accuracy; online, combined with a rule-based state machine, accuracy rose from 78.02% to 93.14%, with a macro F1 of 0.9343 and an inference cycle of 53 ms.
The authors propose Embodied Playground, a benchmark for evaluating how foundation-model agents coordinate physical, strategic, and spatial reasoning across long-horizon tasks. It contains 410 task environments spanning basic manipulation, sequential puzzles, and multi-stage household activities.
The paper proposes the ERR-Env verifiable interaction environment, the ERR-Bench benchmark, and an online reinforcement learning method, Action-Grouped Turn-GSPO (AGT-GSPO), to train the 2B embodied VLM ERR-VLM to learn replanning and recovery after execution errors.
DexGraspFlow proposes a conditional link-level SE(3) flow-matching framework for cross-embodiment dexterous grasp generation. It combines geometric relations with kinematic topology, uses a dual-stream DiT to model link-link coordination and link-object interaction, and provides explicit structural information through attention biases derived from kinematic-tree distances.
S-Agent 2.0 is a spatial reasoning agent that equips an off-the-shelf general coding agent with a command-line perception toolkit, designing the interface around acquiring scene evidence, writing it into persistent scene memory, and interpreting readings in plain spatial language.
The paper proposes Video-Point Model (VPM), a world model that conditions on multi-view video histories and language instructions to jointly predict future videos from arbitrary viewpoints and the 3D trajectories of persistent scene points. The authors add camera-conditioned multi-view processing and a trajectory prediction head to a pretrained world model, and build the VPM-Data simulation corpus and the VP-WorldBench benchmark; they further add robot actions as a token modality in the generator, so that the policy shares the point-supervised backbone.
The paper identifies two types of decode amplification when randomly sampling videos for AI training: sequential reconstruction of unrelated frames, and reconstruction of the reference chains needed for sampled targets. The authors propose dependency-aware sparse decoding, which reconstructs only the requested frames and their necessary references without transcoding the source video; optional offline DDRA encoding constrains each interior frame between two intra-coded anchors, so that reconstruction of each requested target requires no more than three frames. On a large robotics corpus, the method achieves 1.45x the loading throughput of LeRobot v3 at similar reconstruction quality, with 58.7% less media data.
CausalAlign proposes a causal alignment framework that uses linear probes and causal interventions to locate the representation layers for camera ego-motion in video generation models, then aligns the representations at those layers to a 3D foundation model via delta-token space. After fine-tuning Wan2.2 on DL3DV and RealEstate10K, geometric consistency improves by more than 20% over the base model, and on RoboCasa365 mobile manipulation tasks it raises the success rate of the world action model (WAM) by 10 percentage points.
The paper proposes LaLa-World, a language-conditioned latent world model (LWM) that learns jointly from simulation and robot trajectories as well as broader image and video data. Its key design is to describe both numerical actions and visual changes in language, enabling training of a single future predictor across heterogeneous datasets, and to build the predictor on a language model so it can interpret language conditions.
CausGround proposes a non-parametric causal interaction grounding framework that distills interaction experience into action-centered precondition and effect knowledge, and combines it with current observations during planning for backward and forward reasoning.
TARSphere proposes a full-stack framework for robots that evaluates long-horizon embodied agents under physical and social constraints. Its TARSphere-Infra runs robot policies in a physics simulator with embodied NPCs and constrained dialogue, and uses an event-driven finite-state machine to update task states based on completed actions and delivered information; TARSphere-Harness lets a frozen multimodal large language model revise its decisions through multi-level executable skills, observations, hierarchical memory, and execution feedback.
The authors deployed decentralized multi-agent reinforcement learning (MARL) on two physical robots for joint task assignment and navigation in a setting where multiple robots serve multiple production machines. Policies were first trained in high-fidelity simulation in Isaac Lab/Isaac Sim, then transferred zero-shot to the real robots via a self-built ROS 2 deployment pipeline. Paper: https://arxiv.org/abs/2609.14567.
A Binghamton University team presented two papers at 2026 IEEE/RSJ IROS, demonstrating that implanting hidden trigger words in large language models can cause robots to deviate from their task plans and perform dangerous actions.
Session 172 of the RLChina Paper Seminar was livestreamed on Bilibili and Tencent Meeting on October 20, 2026, 16:00-17:00 (UTC+8), hosted by Zhao Hanye, a PhD student at Shanghai Jiao Tong University, with Zhang Wenhao and Cao Jiahang each presenting one of the two studies.
The paper proposes performing 3D scene generation directly in a high-dimensional implicit 3D latent space derived from 2D foundation models (such as DINOv2, SigLIP2, and DA3), using a diffusion Transformer to model within that 3D latent space and achieve 3D-grounded scene generation.
NASA researchers, working with Imperial College London and Nagoya University, have designed a new Mars helicopter rotor blade airfoil called roamx-0201. Through high-fidelity simulations and wind tunnel tests, it could increase the maximum single uninterrupted flight distance from Ingenuity's 0.7 km to up to 18.7 km while carrying scientific instruments.
The PANVIS STAR® neurointerventional robot completed full-workflow validation of five major procedure categories in a porcine model: cerebral angiography and treatment access establishment, aneurysm coil embolization, flow diverter placement, vertebral artery stenting, and carotid artery stent implantation. All procedures were successful, with no device-related complications or vascular injury. The study was published in the Journal of Robotic Surgery.
3D Vision Workshop introduces a CVPR 2026 Highlight work that proposes a framework named 3D-RGBX, which can synthesize thermal, near-infrared, and SAR images aligned at the pixel level with RGB views by running COLMAP only on RGB, without requiring metric depth, pose, intrinsics, or calibration of the X sensor.
Algorized and the ZHAW Zurich University of Applied Sciences Institute of Signal Processing and Wireless Communications (ISC) have launched a two-year research project, Real-Time Human Perception via Edge-AI Wireless Sensor Fusion for Collaborative Robotics, funded by InnoSuisse and starting in September 2026.
The Gao Fei team at Zhejiang University, together with Weifen Zhifei, launched the CrossFly benchmark, building 23 reconstructed real-world scenes and 15,263 complete trajectories, covering three task types: indoor, outdoor, and indoor-outdoor cross-domain, accompanied by a hierarchical VLA baseline, CrossFly-Agent; the full simulation dataset, evaluation protocol, and baseline code are all open source.
A Reddit user proposed designing a tiny RGB + depth recorder for hand-object manipulation demonstrations and compared single-frame model outputs from three depth approaches in the same synthetic scene (a chest-view grasp of a mug): a D405-class stereo camera, the TMF8829 dToF (48 × 32 zones), and a generic iToF (240 × 180).
RoboValue introduces a fine-grained sim-and-real benchmark for unified evaluation of robotic value models, assessing four complementary capability dimensions through shared interfaces: task-state understanding, temporal progress monitoring, failure and recovery reasoning, and value consistency.
This paper systematically studies the mechanisms, scaling behavior, and generalization of physical recursive self-improvement (RSI) in embodied agents, proposing PhysRSI-eval, which quantifies physical RSI across five dimensions: first-attempt performance, efficiency, success rate, budget, and improvement from failure, and evaluates multiple VLMs on 18 manipulation tasks under a fixed interaction budget.
The research team proposes HandsOn, a composable cross-embodiment dexterous manipulation simulation platform and benchmark supporting 5 robotic arms, 6 dexterous hands, and about 200 rigid, articulated, and deformable assets. The benchmark contains 18,000 human demonstrations and 66 human-verified tasks, of which Core-50 is used for demonstration-based learning and execution tests, while Transfer-16 evaluates asset and relation transfer, skill composition, and deformable manipulation without target-task demonstrations.
The paper proposes the collective experience learning (CEL) paradigm, in which deployed robots share execution experience to improve a common policy without collecting additional expert demonstrations or corrective actions. It uses failure-guided experience retrieval fine-tuning (FERT) to retrieve relevant trajectory segments from the fleet's shared successful experience as supervision, and a fault-tolerant asynchronous training system (FATS) to coordinate experience collection, policy optimization, and distribution.
The study proposes the HOLI-WORLD integrated ecosystem, which connects experience collection, world representation, automated data construction, model learning, and independent evaluation. HOLIWORLD-Foundry combines static and dynamic perception to automatically construct world records, and HOLIWORLD-Data organizes approximately 100 self-collected scenes into approximately 80K QA pairs used to train HOLIWORLD-VL-4B; independently human-annotated held-out recordings form HOLIWORLD-Bench.
AdaExploreBench introduces a benchmark that instantiates environmental dynamics as hidden rules, requiring LLM agents to actively acquire evidence and infer the rules, and, under implicit rule drift, to detect that prior knowledge has become stale and re-explore. Experiments on eight leading LLMs show that GPT-6 Sol achieves only 61.9% calibrated accuracy; providing sufficient evidence or removing rule drift improves average performance by close to 10%. The results indicate that agents can reason from valid evidence but struggle to actively acquire evidence and to recognize when evidence becomes invalid.
VLDA proposes modeling task-relevant state transitions with compact Delta variables, replacing VLA's direct observation-to-action mapping and the full future-state prediction of world-action models. It learns Delta under a unified objective via LT-Filter and GT-Anchor, and DDAR injects Delta into the action denoising process as a structured seed. Experiments implemented in the open-source training framework LoongForge show that VLDA improves long-horizon consistency, robustness, and generalization.
The study introduces the AnyGoal benchmark and open-source library to evaluate world models' compositional generalization on goal-shifted task variants. The task design ensures that individual state factors and actions remain within the training support, but their combinations are underrepresented in expert trajectories. Experiments across multiple environments and world model families show that current models consistently fail under these recombination scenarios, and the authors further analyze where failures occur at the dynamics and representation levels.
The paper proposes TokenWorth, a lightweight online memory framework that learns query-relevant memory value from sparse execution feedback and applies it consistently to memory retrieval and bounded memory management. Across three LifelongAgentBench environments and in both local and API-based deployments, TokenWorth achieves the highest task success rate among the evaluated methods while using the fewest output tokens. The code is publicly available at https://github.com/AIGNLAI/TokenWorth.
EgoXperience proposes a complete pipeline covering data collection, automatic annotation, unified modeling, and benchmark evaluation, aimed at jointly reconstructing observer motion, scene geometry, and hand-object interactions from egocentric video. At its core is the automatic annotation engine EgoXX-Engine, which combines geometrically verified object tracking with SLAM-based world alignment to produce the 50-hour dataset EgoX-Data-6M (with second-level task descriptions) and the 1-hour evaluation set EgoX-Bench.
The paper proposes the Stateful Harness framework, which separates world evolution from video rendering, using a frozen video model to render events and write back the world state, in order to support rule-governed interaction with hidden state. The authors build the WorldState-Bench benchmark, containing 50 counterfactual twin pairs that share an initial frame and action but differ in hidden state and required outcome.
The paper proposes InfiniDream, a VLM-driven framework for open-ended procedural 3D world generation. Its core components include Symbolically Constrained Asset Generation.
X Square Robot released the X2Real simulation benchmark for evaluating the performance of embodied models in the real physical world, proposing three design principles: credibility, validity, and objectivity. In testing, X2Real simulation evaluation achieved a 0.96 correlation with real-robot evaluation in task success rate and a 0.99 correlation in progress score; control parameter calibration achieved millimeter-level trajectory consistency, covering embodiments including Franka, ArtiXon Arm-6A, ARX R5, and Quanta X1.
Fréchet Forcing proposes a fine-tuning method that keeps streaming video world models stable during long-horizon generation beyond the training length and lets them continue to follow action commands. The method applies a flow-matching loss where self-generated history meets real data, constrains context fidelity with a Fréchet distance centered on the real clip itself, and enforces visual realism with a Fréchet distance on real frames in a frozen perceptual space. All three constraints come from short video clips, with no teacher model, online judge, or long videos needed.
ScaleNav proposes a unified world-action framework that achieves language-conditioned local navigation through three-stage training: first, instruction-conditioned video pretraining on internet first-person videos to learn how observations evolve as the camera approaches a target; then, aligning predicted executable action chunks using action-labeled simulation data; and finally, jointly fine-tuning all trainable components.