Gamma Intelligence, TeleAI and Unitree have jointly launched LocoManiLab, a unified evaluation framework for humanoid robot mobile manipulation. It covers five types of home spaces, 13 scenario instances and 234 tasks, with the Unitree G1 and Dex3 dexterous hands as the unified test platform.
Anthropic released two models, Claude Fable 5.2 and Claude Mythos 5.2, which share the same underlying model. Fable 5.2 comes with cybersecurity and biosecurity safeguards and is available to all users, while Mythos 5.2 requires an application through a verification program.
AGIBOT has proposed a robot workload measurement framework called Worken, which calculates "effective intelligent working hours" as "number of active robots × autonomous working duration × output efficiency ratio," linking the effective work robots actually complete with autonomous operating time and output quality.
Ant Group's Lingbo on November 10 open-sourced LingBot-Touch, a tactile pretrained foundation model for dexterous manipulation that can encode different sensing signals such as piezoresistive and vision-based tactile signals into unified tactile tokens for use by different dexterous hands and policy models.
Why it matters: The original article provides comparative data on the cross-sensor tactile pretrained model across 26 real-robot tasks, which can be used to assess the actual impact of tactile representations on policy success rates.
NVIDIA launched the open world action model GR00T 2 and the open-source robot learning framework Isaac Lab 3.0 at CoRL. GR00T 2 was trained on 100,000 hours of NVIDIA open-source first-person data; with only one demonstration per task, it reached an average success rate of 71% across 50 tasks, rising to 95% with five demonstrations, and without task-specific demonstrations it completed at least once each of 40 of the 50 tasks.
Why it matters: GR00T 2's few-shot success rates and Isaac Lab 3.0's training throughput gains provide the current baseline for embodied models and the simulation training stack.
Flex-RSI is a benchmark for evaluating agents' recursive self-improvement capability under a fixed budget, covering perception, reasoning, robot action, 3D reconstruction, visual creation, and infrastructure tasks. In each round, an executor improves the task solution, which is kept only if it scores higher on held-out validation data; an improver then revises the harness, including the executor and itself. Final solutions are scored on hidden test data.
The paper proposes agentic manipulation policies (AMP), in which a generalist AI model directly controls a robotic arm and adjusts its actions as the task progresses. The authors' reference implementation, AMP1, runs unchanged across single-arm, bimanual, and mobile manipulation.
Why it matters: The paper reports zero-shot success rates and real-world trial data for a generalist AI model directly controlling a robotic arm, which can be compared with specialist policies such as VLA and WAM.
The authors propose the Seeing What Matters benchmark to evaluate task-conditioned actionable spatial belief in simulated tabletop manipulation, covering 860 frozen evaluation cases across three scenarios—coffee workstation, drawer/cabinet, and microwave—which expand to more than 115,000 evaluation episodes.
The research proposes having GPT-6 Astra read a robot's base policy weights and edit only 65 rows of token embeddings, raising the success rate on π0.5's 765 held-out paraphrased instructions from 86.3% to 92.0% (48 repairs, 4 regressions, paired p=1.31×10⁻¹⁰), all without gradient updates or new demonstrations.
The study proposes a planner-verified counterfactual (CF) test: holding the grasp pose fixed, a solid obstacle is inserted into the approach corridor and its points are added to the model input, after which the planner re-derives the label, flipping it from feasible to infeasible. Across eight predictors spanning point clouds, voxels, depth, and object tokens, CF-pair AUROC falls from 0.920-0.996 in natural scenes to near chance (0.470-0.504), and this flip holds across four planners, two robotic arms, and a real-object obstacle.
Unitree Technology shipped over 5,500 pure bipedal humanoid robots in 2025, with humanoid robot revenue of 868 million yuan. After listing on the STAR Market in August 2026, its market value has retraced more than 55% from its peak. Daily rental rates for humanoid robots fell from about 20,000 yuan in early 2025 to the thousand-yuan level in 2026. In the first half of 2026, global shipments exceeded 22,000 units, but only about 4,000 entered factories; entertainment and commercial performances plus science education and data collection together accounted for over 60%.
Nuoyin Intelligence says KnowinAgent achieved a team self-tested composite score of 87.60 and an overall success rate of 82.83% across all 54 task variants in RoboDojo, once again ranking TOP1 on the leaderboard.
A user has released free comparison content of the Unitree G1 EDU+ and the Basic G1, covering the unboxing, hardware differences, walking performance, and App features. The video is a side-by-side demonstration, and the creator is soliciting the community's views on the differences between the two versions on Reddit.
A team from Northwestern, CMU, UNC Chapel Hill and Amazon, in the paper "WOVEN: Weaving Visual World Modeling into Multimodal LLMs," defines world modeling as estimating the state transition p(s′|s,a) and builds WOVEN, a training and evaluation base of 36K questions.
Why it matters: The paper breaks world modeling into independently intervenable state-transition tasks and provides reusable rules for selecting training data, offering reference value for multimodal model training.
miHoYo, together with Tsinghua University, the University of Hong Kong, the National University of Singapore, and other institutions, has open-sourced the GameScaling series of general game models, covering four size tiers—0.8B, 2B, 9B, and 27B—for a total of eight models. The weights and inference code are released under the Apache 2.0 license.
A team from Xiamen University and Nanyang Technological University proposed RoTA-Bench, which uses interactive 3D-to-2D projection alignment to evaluate whether vision-language models can precisely adjust an object's 3D pose based on visual feedback. The benchmark contains 674 tasks, of which 100 form a frozen test set; models must choose among 24 discrete rotation actions so that the projection reaches an IoU of 0.90 with the target contour.
Tristan, co-founder of YC F26-incubated company Asim, said the team is building a standardized benchmark for physical intelligence in humanoid robots, with a focus on evaluating dexterous manipulation capabilities. The benchmark uses repeatable scenarios to examine spatial reasoning, grasp adaptation, coordination, and how robots respond to physical interaction, with the goal of pinpointing where capabilities fail and providing reproducible failure cases. The team also noted the need to address comparability issues caused by differences in hardware, sensors, and control interfaces.
Jiyuan Ludong has released NeoHorse-1.5-Flash Preview, an agent-task-oriented model, and opened it for internal testing. The model is built on Qwen3.8-Flash through recursive self-improvement iterations, with 125 billion main model parameters and 6 billion activated parameters.
AGIBOT has released GO-3 Preview, an embodied foundation model built on its self-developed WRAM (World Reasoning Action Model) architecture that unifies task understanding, future state prediction, and action generation. It ranks first on three embodied AI benchmarks: RoboDojo, RoboTwin, and RoboColiseum, and is planned for full open source release by the end of 2026.
Why it matters: AGIBOT has disclosed the architecture and three-stage training design of GO-3 Preview, which can be compared against its technical roadmap evolution from GO-1 to GO-3.
After GPT-6 Astra's release, general large models can directly control robots to perform tasks; embodied data orders are rebounding, the value of body companies is being reassessed, and model companies building embodied brains now face competition on the same field as foundation model giants. In NVIDIA testing, Astra achieved a 66.3% success rate in zero-shot direct action output, higher than π0.5, but average single-inference time was 18.56 seconds, and one successful action consumes about 160,000 tokens, equivalent to $1.96. The industry currently prefers having general models handle planning and the semantic layer, while embodied models handle millisecond-level motion control and the execution layer.
The authors propose VABAR, a Variational Automation Benchmark for agentic robotics, built on open-source tools and the Newton physics engine for simulated manipulation evaluation. The benchmark covers six manipulation task classes: tool sorting, tool hanging, gear meshing, syringe packing, cable insertion, and cable routing. It includes standardized task variants, difficulty levels, success metrics, 33 generalizable skills, and human-written programmatic reference solutions. The paper also presents an initial baseline evaluation based on Graph-as-Policy (GaP) and Isaac Lab Arena.
The authors propose Compounding Physical RSI, a self-evolving policy framework that uses automated capability-gap assessment to locate the task stage where a failure occurs, refines policies with behavior-regularized offline reinforcement learning, and builds candidates in weight space to preserve existing skills.
The study proposes a failure-boundary profile evaluation method in which each scenario family varies only one measurable factor (such as corridor width or the time gap in oncoming traffic), and each variant is labeled from measurement as one where the agent should proceed or one where it should stop. Along this axis, two curves are estimated for each agent: commitment (the probability of attempting the passage) and competence (the probability of passing without contact). The method locates where each curve crosses 0.5 and gives intervals that account for the simulator's run-to-run variance, while also using certified labels to derive a decision accuracy metric.
The study proposes RACE, a robot policy checkpoint scoring metric that requires no real-world rollout. It uses a fixed set of held-out successful trajectories to calibrate action deviations against state-conditional thresholds, then applies a running product to reduce the weight of later reference predictions after large deviations. In simulation, RACE outperformed baselines across different validation sources, while some baselines degraded noticeably when the validation set changed; on a real robot, its rankings were consistent with deployment success rates across the evaluated policy backbones and tasks of varying difficulty.
RoboColosseum is a benchmark for testing policies on real robots, covering 150+ scenarios and hundreds of actual robots, with all rollouts (including failures) published on Hugging Face.
Google and Mercor released the Physical-Engineering's Last Exam (PLE) benchmark, which evaluates engineering agents on 85 mechanical, electrical, civil, and structural engineering tasks.
The China Academy of Information and Communications Technology (CAICT) and Tsinghua University's Department of Electronic Engineering jointly released the Embodied Intelligence Development Report (2026), stating that the industry is "hot on the surface but cold underneath," that a shakeout has begun, and that model approaches are far from converged. The report says domestic embodied intelligence financing totaled RMB 93.5 billion in the first half of 2026, exceeding the RMB 73.5 billion for all of 2025, while global humanoid robot shipments exceeded 22,000 units; however, education and research institutions accounted for about 55.5% of domestic winning bids, while industrial and technology enterprises accounted for only about 21.1%.
The study proposes Route Anything, a unified framework for studying model and policy routing in general-purpose AI systems, and establishes an optimal routing theory under quality-cost trade-offs. The authors construct an omni-modal trajectory matrix covering 25 benchmarks across language, image, video, speech, coding and workplace agents, and robotic manipulation, with 452 candidate configurations and 374,401 scoring runs.
The article notes that humanoid robots are moving from demonstrations to mass production, but the manipulation success rate drops sharply from 89.4% in simulation environments to 12% in real home scenarios, and reliability must be supplemented by a testing system. Integrated joint modules account for about 40% of the BOM cost of the complete machine, and GB/T 43200-2023 has provided a testing framework in four categories: motor performance, control performance, positioning accuracy, and mechanical performance. NI uses a cRIO joint test bench and a hardware-in-the-loop simulation solution based on Isaac Sim, PXI, and RDMA to try to insert an HIL stage between pure simulation and real machines, narrowing the gap in migration from simulation to reality.
Tsinghua University and UC Berkeley jointly proposed and open-sourced uni-robot-agent, a cross-environment unified embodied agent harness for connecting tool interfaces across different robots and simulation environments. Based on this framework, Qwen3.8-Flash-Next scored 31.39 on the RoboDojo benchmark, exceeding the publicly reported 28.97 for GPT-6 Astra, at about 1/20 of Astra's cost.
The Xbotics Embodied Intelligence Lab built a closed-loop simulation pipeline using the AgiBot GE-Sim 2.0 world model and the OpenPI π0.5 policy. In testing, the world model generated about 4.33 frames per second, the pure-policy closed loop ran at about 3.20 frames per second, and after adding MiMo online scoring, the rate dropped to about 0.88 frames per second.
Why it matters: The author uses a four-stage test to break down the frame-rate bottlenecks and causes of reward hacking in world-model closed-loop simulation, which can be transferred to the design of VLA evaluation workflows.
Official weights of InsertAnything's force-conditioned insertion policy were evaluated in Isaac Lab: at 0.1 mm clearance, baseline success rates for two hole types were 96–100%. With a 5 mm initial hand offset, success rates were unaffected, but with a 3 mm hole-position observation deviation, the hexagonal hole success rate dropped to 60%.
The authors propose SimuGuard, an auditing plugin for LIBERO, RoboTwin, RoboCasa, and RoboDojo that identifies failures caused by the simulator rather than the policy. On RoboTwin, the authors trace these failures to the physics engine's collision detection: when geometry is 1 to 2 centimeters apart, the engine reports a penetration of about 2 centimeters, and the solver ejects the object at several meters per second because the distance algorithm discards triangles below a fixed area threshold.
Why it matters: The original traces failures in simulation benchmarks to a physics engine collision detection constant and provides a before-and-after success rate comparison, which can serve as a reference for teams doing simulation evaluation.
The authors propose the MolmoSpaces-CommonSense benchmark, which builds a reasoning taxonomy grounded in developmental cognitive science and examines commonsense reasoning about occlusion, containment, support relations, semantically correct grasp points, spatial affordances, and hidden objects.
The research team used a five-fingered robotic hand to grasp and insert RJ45 connectors, evaluating 281 imitation-learning policies spanning seven policy families in approximately 1,700 real-robot trials. Controlled ablations showed that observation and action representations had a greater impact than conventional optimizer and sampling choices. For point-cloud policies, region-of-interest cropping, representation in the robot-base frame, and normalization within the point encoder each improved success by 70 to 100 percentage points. For the vision-language-action policy, switching to absolute action prediction improved insertion success by up to 70 percentage points.
Why it matters: Approximately 1,700 real-robot trials compared 281 imitation-learning policies, indicating that observation and action representations determine real-world success more than optimizer and sampling choices.
RoboColosseum is a distributed evaluation framework that evaluates general-purpose robot policies over a network of locally operated robots in multiple locations, using unified policy serving and standardized protocols.
Why it matters: This distributed evaluation framework used 5,699 paired trials to compare VLA policies with frontier agents, revealing inconsistent rankings across embodiments and a ranking reversal after fine-tuning.
Xinhua News Agency's China Economic Information Service and the Shanghai Industrial Innovation Center of the China Academy of Information and Communications Technology jointly released the "2026 Embodied Intelligence Industry Innovation Report," which proposes that in 2026 embodied intelligence will move from development demos to routine operation in real-world scenarios, entering a deployment growth phase. The report documents that 8 AgiBot Genie G2 units accumulated more than 64 hours of operation over 6 days of testing at Longcheer Technology's Nanchang factory, completing 64,828 tasks with a comprehensive task success rate of 99.99%, and notes that task intelligence remains slightly below L3, with bottlenecks including non-interoperable data standards, a lack of calibration for data credibility, and ownership allocation still to be improved.
A joint team from Fudan University's Institute of Trustworthy Embodied Intelligence and Shanghai Xinzhi Embodied Intelligence Technology Co., Ltd. won first place in the quality inspection category at the 2026 International Embodied Intelligence Skills Competition. The student team, working with the National and Local Center, used the Linglong H wheeled-arm robot and the Linglong L bipedal robot to win two first places, one third place, and one fourth place across three categories: loading and unloading transport, firefighting, and precision assembly. In the firefighting category, the Linglong L bipedal robot completed fire detection, fire suppression, and evacuation without human intervention.
Xingyan Intelligence has launched Auriga, a general embodied Harness framework. Under a zero-shot protocol that involves no task-specific pre-tuning, no reliance on expert demonstrations, single-attempt execution, and no environment resets, it achieved a 57.74% average task success rate and an average process score of 64.00 across 42 tasks and 6,300 test episodes on RoboDojo, using GPT-6 Astra as the base model. It leads the leaderboard on both metrics.
Xinzhiyuan reported on the SpatialScout benchmark, which gives a camera to an AI and places it in previously unseen virtual rooms, where the model autonomously decides when to turn, move, and capture images, and submits the 3D coordinates of target objects. The formal evaluation selects 80 indoor scenes and 3,140 target instances from iTHOR, RoboTHOR, ArchitecTHOR, ReplicaCAD, and ProcTHOR, with at most 100 interaction rounds per task; scoring considers both mean IoU and the proportion of IoU above 0%, 50%, and 90%.