AgenticRoboRL proposes a framework for post-training robot planners across different robot execution harnesses. The training recipe combines supervised fine-tuning with multi-turn reinforcement learning, and harness-specific adapters connect planner interaction records and task outcomes to a unified RL training pipeline while keeping execution components fixed.
Renmin University of China's GeWu-Lab and the Beijing Humanoid Innovation Center will livestream a presentation of the Robo-ValueRL millimeter-level VLA reinforcement learning solution on October 21, 2026 at 19:00 Beijing time, with talks by Hu Di and Xia Wenke.
A research team proposes RoboJudge-Bench to evaluate how well multimodal large models perform as judges of embodied video generation, covering two dimensions—physical plausibility and instruction consistency—and comprising 800 human-annotated videos from 8 source corpora and 10 video generators.
Professor He Jiong's team at Seoul National University's AIDAS Lab proposed Dynin-Robotics, a robot foundation model that unifies language, images, video, and robot actions in a single model, generating and selecting actions by predicting future scene images after an action; the paper has been accepted by NeurIPS 2026.
The HumanPlus lab at Xiamen University, together with researchers from MIT and CMU, has launched CLOTHO, a full-body motion capture system that integrates 10 IMUs into loose clothing. The paper was accepted to the SIGGRAPH Asia 2026 journal track and will be published in ACM Transactions on Graphics.
WorldOracle, a world-model evaluation platform jointly developed by researchers from AgiBot, NVIDIA, and several universities, is now live and open for evaluation submissions. The platform uses 335 real manipulation trajectories, with about 40% of interaction segments containing failed or suboptimal operations, testing model prediction against real execution outcomes. Testing of 7 world models found that while some models can reproduce robotic-arm motions relatively accurately, their predictions of object states still show considerable deviation, and performance drops markedly on long-horizon prediction.
A research team from Central South University, the Beijing Academy of Artificial Intelligence, Hong Kong Polytechnic University, Beijing University of Posts and Telecommunications, and other institutions jointly released a 57-page systematic review of system-level agent distillation. The paper is titled Distilling Agentic Systems: A Roadmap across Models, Artifacts, and Harnesses, and is available at https://arxiv.org/abs/2609.36630.
X Square Robot has released X2Real, an evolvable simulation evaluation benchmark. Its simulated evaluation and real-robot evaluation achieve a task success rate correlation of 0.96 and a progress score correlation of 0.99.
Why it matters: It provides correlation data between simulated and real-robot evaluation along with a unified integration method, allowing readers to judge whether simulated evaluation can replace part of real-robot testing.
SenseTime SenseNova, together with Daxiao Robotics, Ropedia, Nanyang Technological University Singapore, and the University of Tokyo, released "Has GPT-6-Astra Achieved Embodied Intelligence?"
An MIT team published research in Advanced Functional Materials describing a paper-thin swimming robot driven by a single layer of living muscle cells, which can control direction and speed through flashing light and swim through watery mazes.
On October 16, WorldOracle, an action-driven world-model evaluation platform jointly completed by researchers from AgiBot, NVIDIA, Princeton University, the University of Texas at Austin, and the University of Toronto, was officially released, opening evaluation submissions to developers.
Why it matters: The evaluation compares world-model predictions frame by frame against real execution records, revealing a common problem in which robotic arm motion is replicated but object states are systematically hallucinated, providing concrete reference points for judging model usability.
A research team tested letting a coding agent itself act as a closed-loop robot policy: the agent receives RGB observations and robot state, writes and executes Python control programs, and revises the next program based on actual results, with no learned motion policy at any point.
A team at Delft University of Technology in the Netherlands published the autonomous flight system MonoRace in Science Robotics; at the 2025 Abu Dhabi drone racing competition, it reached a top speed of 100 km/h and defeated multiple AI teams and 3 human world champions.
RealtimeWAM introduces an efficient variant of the World Action Model (WAM) that uses one-step action generation and asynchronous inference to overcome two major efficiency bottlenecks: multi-step denoising and serial waiting between experts.
GroundAnything is a 4-billion-parameter visual grounding foundation model that combines parallel decoding with precise object localization through blockwise denoising. Its training integrates pretraining on public datasets, direct conversion from autoregressive to diffusion, supervised fine-tuning, and GRPO-based reinforcement post-training.
Metro-WM proposes a hierarchical planning framework that, instead of generating unconstrained latent sub-goals, retrieves real states from observed frames in offline expert demonstrations or random-action trajectories as sub-goals, and connects frames from different episodes into paths toward the goal. Experiments show its long-horizon success rate is up to 37.33 percentage points higher than the next-best hierarchical method, up to 10.9 times faster, uses 13-56 times less offline computation, and requires less tuning. The authors also note that leading macro planners frequently output physically unrealizable sub-goals.
EgoTac is a generalizable model that predicts rich tactile information directly from egocentric human videos. It was trained on more than 5.7 million image-tactile pairs covering both continuous force measurements and binary contact. Experiments show an average force error below 0.06N for in-domain prediction, and on out-of-domain contact prediction benchmarks it consistently outperforms the current state-of-the-art contact estimator. It also enables zero-shot prediction on unconstrained real-world videos; scaling analysis indicates that both data diversity and data volume steadily improve performance. The work was accepted to NeurIPS 2026.
The paper proposes CorrelationFlow, which reformulates LiDAR scene flow estimation as a correlation maximization problem between Bird's-Eye View (BEV) images, replacing learnable parameters with a chain of image processing operations. This training-free method won second place in the unsupervised track of the Argoverse 2 Scene Flow Challenge (CVPR 2026). The test set covers five datasets and a variety of vehicle platforms and environments, and results show its performance can match or even surpass learning-based methods. The source code will be open-sourced.
NEXUS is a modular neuro-symbolic framework that jointly maintains a PDDL action domain model and an LTL safety specification: execution failures are used to repair action schemas, while detected hazards are converted into reusable LTL constraints through hierarchical semantic and translation sampling with equivalence consensus.
An ECCV 2026 accepted paper proposes UniWM, a memory-augmented world model that unifies egocentric visual foresight and planning within a single multimodal autoregressive backbone, grounding action selection in visually imagined outcomes and tightly aligning prediction with control.
MAIA is a multimodal framework for anticipating anomalies in contact-rich manipulation. Its vision-action-force world model learns normal interaction dynamics and predicts the visual and force responses of planned actions, while a lightweight anomaly head predicts impending interaction anomalies.
The study proposes a two-stage deployable linkage that self-locks through its own geometry and is integrated into a commercial robot vacuum to support pushing-type tidying tasks. The mechanism uses a single actuator to deploy and lock, a passive trigger to release the lock for retraction, and a second actuator to adjust deployment height only; at full deployment, an internal three-link subchain approaches a collinear, near-singular configuration in which the transmission ratio tends toward zero, so loads produce only a very small residual back-driving torque.
WorldOracle is a benchmark for action-conditioned world models (ACWMs) that uses identical initial observations and low-level action sequences to test whether predicted rollouts can reproduce real-world physical consequences and task-state transitions. The evaluation covers four dimensions: physical state correctness, interaction spatial alignment, key interaction area fidelity, and multi-view action coherence, and is paired with long, multi-segment real-world data containing both successful and failed executions. Experiments on seven representative ACWMs reveal different models' strengths and failure modes. The dataset, evaluation code, and leaderboard are publicly available.
The study uses a frozen frontier vision-language model, GPT-6-Astra, directly as the robot policy and tests it on RoboMME, a benchmark of 16 memory-dependent manipulation tasks. The model keeps the entire episode context throughout and directly outputs each end-effector waypoint and gripper command, with no training specific to this benchmark.
Why it matters: The paper provides complete comparative data on a frozen frontier VLM used directly as a policy on a memory-dependent manipulation benchmark, which can be used to assess how much of the embodied stack a general-purpose model can take over.
An evaluation conducted 864 rollouts of three world-action models across four RoboLab scenes, holding initial states and task goals fixed, and found that describing spatial relations using the reference object rather than the object to be moved lowered the stable placement success rate by an average of 12.2 percentage points. An independent test of the policies' language understanding identified relation reversals, and adding images disrupted otherwise correct text-only interpretations. Rollout analysis also distinguished spatial placement from instruction following: the stable placement rate fell from 43.2% at any time to 21.8% at episode end.
The paper proposes an explicit analytic robot self-model that requires neither deformation networks nor URDF/CAD templates. From multi-view images and joint readings, it obtains appearance and rigid link assignments, and independently identifies metric screw axes using sparse end-effector poses, yielding closed-form forward kinematics and an analytic Jacobian that can be extrapolated.
AnyRank proposes a sequential method for multi-policy ranking that adaptively allocates trials to policies that are hard to distinguish, provides a (partial) ranking at any stopping time, and bounds the probability of an incorrect ranking. The method can optionally use prior information (such as expert intuition, a small set of held-out hardware evaluations, or policy rollouts from a simulator or an action-conditioned video world model) to improve hardware evaluation efficiency, and inaccurate priors do not break its statistical guarantees.
An arXiv paper presents RoboHarness, a robot harness that gives LLM agents a visual-geometric control panel so they can directly understand and invoke embodied tasks. The authors argue that coding agents have the potential to achieve generalist embodied manipulation without the route of large-scale VLA training. Experiments show that, with RoboHarness, Qwen3.8-Flash-Next outperforms state-of-the-art VLA models on BEHAVIOR Challenge 2025 tasks, and the implementation is open source.
The research team proposes DexBioLab, a simulation benchmark and framework for dexterous manipulation in real biological laboratory environments, comprising 14 tasks across 4 robot embodiments, covering functional tool use, precision manipulation, language-conditioned visual grounding, and instrument operation.
The paper proposes using Conditional Flow Matching to reformulate long-horizon goal prediction in open-environment active 3D mapping as conditional multimodal anchor generation, replacing traditional single-point prediction.
A tendon-driven bionic jellyfish robot achieves 150° bending through soft actuators that combine a flexible substrate with discrete constraints, with tendon displacement and bending angle approximately linear. Eight actuators driven by four servos enable stable swimming, attitude adjustment, and self-righting. A reinforcement learning controller trained on this linear actuation characteristic achieved closed-loop depth regulation in both simulation and physical experiments.
This study proposes an interdisciplinary process for benchmarking service robots in public everyday environments, addressing the gap between laboratory metrics and real-world operation, with evaluation conducted at technology readiness levels 6-7. Over the three years 2023-2025, seven benchmarking events (one consensus workshop and six on-site evaluations) assessed three robots across three use cases: outdoor park cleaning, pedestrian underpass cleaning, and interactive assistance in a public library.
The CoRL 2026 paper proposes OpenDexGrasp for open-vocabulary task-oriented dexterous grasp generation, enabling a robot to infer functional intent from free-form language and, combining multi-view visual observations with object geometry, generate executable high-degree-of-freedom grasps.
The paper proposes RoboReel, a unified benchmark for evaluating models that learn policies from human videos (Learning from Observation), featuring real human demonstration videos, simulated robot trajectories, and evaluation environments for ten manipulation tasks.
This study analyzes reinforcement learning policies for legged locomotion using the effective rank of the network Jacobian, finding that conditioning on gait phase exposes representation structures invisible from the global rank. Architectures with observation normalization and residual connections allocate more than one additional effective-rank dimension to swing-dominant states than vanilla MLPs do. Based on this, the proposed diagnostic method reduces joint jitter by up to about 2x in simulation and on a physical Spot, and the paper has been accepted to CoRL 2026.
The paper proposes DriveDegrade, an image-degradation tolerance benchmark for camera-only end-to-end (E2E) driving. It injects 16 corruption types at 5 severity levels online in the image loader, uses a single operator to cover 15 policies, and evaluates open-loop planning on nuScenes and NAVSIM, with a CARLA closed-loop anchor.
ADEPT is a large-scale reinforcement learning framework that first pre-trains a dexterous policy on a generic object-reposing task, then post-trains downstream policies with this prior, enabling long-horizon tasks to be solved directly from raw visual-tactile perception across high-DoF embodiments. To prevent naive fine-tuning from degrading zero-shot reposing ability, the authors propose a stable post-training scheme combining behavior-cloning distillation, critic warm-up, and conservative on-policy updates, and introduce a joint-space Geometric Fabric that coordinates the policy with the robot.
Cyclops is a framework that synthesizes sparse Non-Repetitive Scanning LiDAR (NRS-LiDAR) intensity data into RGB video for all-weather perception tasks without a camera. Its method first uses a frozen pre-trained densification module to convert sparse intensity projections into dense representations, then learns a velocity field through Latent Bridge Matching (LBM) to transport intensity latents to the RGB distribution within a few ODE integration steps; it also injects prior-frame context via temporal attention to suppress inter-frame flickering, and models the velocity field as a policy optimized by a differentiable terminal reward.
MANGO-Grasp proposes an anisotropic interaction framework that represents objects with geometry-oriented 3D Gaussian primitives to preserve local surface geometry, encodes hands as pretrained descriptors that capture both morphology and kinematics, and models hand-object contact through anisotropic Mahalanobis fields.
The paper investigates why vision-language-action (VLA) models fail in contact-rich manipulation tasks that require precise physical interaction, and proposes two failure modes: precision failures (policies struggle to produce the small, precise corrective actions needed during contact) and force failures (policies fail to respond to contact forces).