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.
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.
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.
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.
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.
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 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.
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.
WireCraft is a simulation benchmark for industrial deformable linear object (DLO) manipulation, covering three task families: connector insertion, clip routing, and channel seating. It supports two physics models, articulated and deformable, and the released trajectories come from both simulation and a real UR5 robotic arm.
SidewalkBench is a benchmark for visual navigation on urban sidewalks built on NVIDIA Isaac Sim. It provides GPU-accelerated, high-fidelity sidewalk environments, including procedurally generated and real-world scanned scenes, and adds event-driven pedestrian reactive behaviors. The authors evaluated 9 visual navigation models on 330 unit-test scenarios, 800 pedestrian-reactive scenarios, and 105 long-horizon scenarios. They found that pedestrian interaction and long-horizon robustness remain key bottlenecks for existing models, and that scaling up sidewalk training with synthetic data is a viable direction.
This study presents the first systematic study of confidence calibration for vision-language-action (VLA) foundation models, establishes a baseline for VLA confidence estimation, examines the relationship between task success rate and calibration error as well as how calibration changes over time, and proposes two lightweight techniques, prompt ensembles and action-wise Platt scaling, to correct the observed miscalibration. The paper is 34 pages long with 7 tables and 22 figures, and has been published in TMLR as a final version.
AWAM (Active World-Action Model) introduces a virtual active-perception paradigm that actively plans observation positions and, before execution, imagines what will be seen along the planned trajectory. It treats the predicted task-related virtual camera motion as a structured query into a geometry-conditioned world model, yielding a reusable spatial prior that is compressed into a fixed memory and fused with live observations through cross-attention for diffusion-based action prediction.
The paper proposes an Interventional Memory Evaluation Protocol (IMEP) for vision-language-action (VLA) systems, used to determine whether memory truly influences robot decision-making, covering four questions: the net benefit of memory, the conditions under which it takes effect, changes to task outcomes, and whether the effects are specific to task-relevant facts.
LY-GWM-RoboCasa-Human300 combines a Human300-fine-tuned GR00T-N1.5 action generator, an action-conditioned graph dynamics model, and a binary state-reward scorer; at inference it samples 4 action chunks and executes the candidate with the highest predicted-state reward.
The NeurIPS 2026 paper NestedVLA proposes using a nested hypernetwork to generate task-adaptive action-expert parameters for a target task, combining them with a frozen base VLA. This replaces parameter aggregation with parameter generation, mitigating interference in multi-task training.
SimpleVLN is a vision-language navigation (VLN) framework that preserves the native architecture, achieving success rates of 76.7% and 80.1% on the R2R-CE and RxR-CE splits, respectively, significantly outperforming existing methods.
The paper proposes Embodiment-Classifier Guidance, which trains a lightweight embodiment classifier and uses the classifier's gradient during denoising steps to guide a frozen π0 policy, enabling transfer to new robot embodiments without additional demonstrations or parameter updates.
The SenseNova joint research team released a report evaluating GPT-6-Astra's embodied capabilities in robotic manipulation and navigation. Under a unified protocol, the team evaluated it on 12 representative benchmarks spanning tabletop manipulation, humanoids, mobile platforms, and dexterous hands, allowing in-context learning (ICL) from training-set demonstrations and substantially increasing the number of execution steps, achieving what the paper calls unprecedented performance.
SpatialScout is a new benchmark for evaluating vision-language model (VLM) agents' ability to actively build object-level 3D maps. It contains 80 diverse indoor scenes and 3,140 target objects. The task requires agents to control a camera to explore scenes and estimate object centers in a shared coordinate frame, coordinating viewpoint selection, cross-view association, geometric estimation, and map revision within a limited interaction budget.
The paper proposes Ropedia W0, an egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. The research builds the Xp-Data dataset, containing 10,000 hours of egocentric human data and 3,000 hours of robot data, with unified multimodal annotations and quality control; camera poses encode locomotion and projected skeletons encode manipulation, with robot skeletons combining each embodiment's forward kinematics and manual keypoint annotations.
ACE-Brain-0.6 combines robot manipulation recordings with complementary visual data from multiple sources through a unified data pipeline; after automatic annotation, the data is converted into verifiable grounding supervision, and Qwen3.5-9B is fine-tuned on about 1.5 million samples to produce the model. It outperforms GPT-6 Astra on four of seven public grounding benchmarks, with an overall average score of 75.41 versus Astra's 73.68. Under a fixed fine-tuning setup, aggregate grounding performance improves steadily as the dataset expands to cover more grounding scenarios.
Researchers propose PVBench, a physical video benchmark that quantitatively measures the physical plausibility of generated videos across 46 dimensions in seven domains, covering instruction and semantic fidelity, behavioral and causal consistency, perceptual and temporal quality, geometric and spatial consistency, material properties and field states, kinematic and spatiotemporal persistence, and mechanics and conservation laws.
StartLux-Decision-v1.1-27B Instant, from Chinese AI startup StartLux (Yuandian Xinghui), ranked first worldwide on the Hugging Face Decision Index 0.3.2 with a score of 69.1, placing it at the top among 173 publicly reproducible models. It leads the second-place model by nearly 3 points and is the only model on the list to exceed 69 points. After releasing its first-generation model in late September, StartLux set a new best result in just over ten days, maintaining its position as the global SOTA in decision models.
The paper proposes BiDex-100, a simulation benchmark for dual-arm dexterous manipulation comprising 106 tasks organized along five capability axes: generalization, instruction following, limited demonstrations, long horizon, and memory. Each task is paired with an out-of-domain variant featuring controlled visual and spatial changes to assess robustness.
A G1 shooter based on motion-prior reinforcement learning scored on all 60 penalty kicks, and a goalkeeper using historical encoding saved 85% of shots. However, when the two models were placed in the official adversarial scenario, they scored zero goals across 20 matches; the article uses frame-by-frame data to identify a two-layer distribution-boundary mechanism.
Manifold AI has launched WorldScape R²SI, a recursive self-evolution framework for robots, which has taken the top spot on the overall RoboDojo robot manipulation benchmark leaderboard, surpassing models such as GPT-6-Astra and NVIDIA GR00T-N1.7.
The Surface Laptop Ultra is equipped with the RTX Spark N1X chip, offering up to a 20-core Grace CPU, a Blackwell GPU with 6144 CUDA cores, and 128GB of unified memory. It is rated at up to 1 PFLOPS of AI compute at FP4 precision. Pre-orders open on October 7 and it goes on sale on October 16, with a starting price of 21988 yuan and 48388 yuan for the 128GB top configuration.
Scale AI, in partnership with Elorian, released a new benchmark called Humanity's Sixth Sense (HSS) to test intuitive visual reasoning in multimodal models: 20 human participants scored 93.1%, while top-ranked GPT-6-astra reached only 53.6% even at maximum reasoning effort, and the median across 25 models was 30.9%.
physim-gate v0.3.1 fixes three issues found in the v0.3.0 review: comparisons could silently produce wrong verdicts or exit codes. When trial ids are shared, a trial reported by only one side is recorded as a missing value (physim.failure_rate counts it as a failure; fully disjoint trial ids are still compared as independent samples). Duplicate keys in gate configs, suites, run results, and LeRobot output are rejected, with the file name identified.
The RL Alpha Team formed by the Zhuoyide technical team took first place in Task 3 "Assisted Living and Feeding" among nearly 100 teams in EBiM Challenge 2026, the world's first intercontinental dual-arm manipulation benchmark.
Google has released Gemini 4 (Argon), positioned as a general physical reasoning model focused on interaction with the physical world, covering the full robotics lifecycle from control execution and simulation reconstruction to data annotation and self-improvement. On the ERQA embodied reasoning benchmark, Argon scored 87%, surpassing Claude Opus 5.5 (82.3%) and GPT-6 Astra (81%); in the AA-Omniscience test, its hallucination rate was 15%, lower than the 51% and 59% of the latter two.
Why it matters: The article breaks down the model's specific uses and measured figures across the robotics R&D chain in four layers: control, simulation, data quality inspection, and self-improvement.
Researchers propose HexaShell, a system in which a coding agent built on a vision-language model drives a dual-arm robot through a single command line; the framework only converts the agent's commands into robot motion. After each failure, the agent reads the log, identifies the cause, and writes or fixes small tool programs. The tools use only the robot's own observations, determining the relative positions of objects through triangulation and, when necessary, moving a wrist camera to obtain a second viewpoint.
Researchers propose OfficeTown, a professional organization simulation framework initialized from a natural-language description of an organization; members coordinate work, delegate tasks to assistant agents, and review deliverables, while artifacts, decisions, and unresolved matters continuously shape subsequent work.
UBTech has released its self-developed unified multimodal world model Thinker-WM, which ranked first on all three benchmarks: WorldOlympiad (overall score 0.740), EWMBench (total score 4.66, surpassing Qwen-RobotWorld and Sora2), and WorldArena1.0 Track-1.
MIST separates "executing the change" from "remembering the change" in video world model evaluation, locating the bottleneck in the model's ability to set up a scene but failure to carry out the instructed change. Based on this, the proposed M3-Agent (Make the Missing Move) recursively revises the details of the input instruction, enabling a frozen model to successfully execute the instruction, with the reveal score more than tripling relative to the baseline.
LoopWAM, a world action model developed by Origin Force Robotics in collaboration with Tsinghua University and Jihuan Era, has topped the RoboChallenge Table30 V2 leaderboard, ranking first in both success rate at 43.33% and process score at 56.80.
IDC's "Worldwide Quadruped Robot Market Shares, 1H26" shows that in the first half of 2026, the global quadruped robot market exceeded $400 million, up 95.4% year over year, with shipments exceeding 49,000 units. Education and research and home consumer use contributed about 60% of shipments but accounted for less than one-third of revenue.
Shanghai AI Laboratory's embodied evaluation team used EBench to evaluate GPT-6-Astra as a robot policy directly, without task-specific post-training, across 26 tasks and 510 total executions. Its average task success rate of 46.7% ranked third, and its average task score of 0.654 ranked second; OpenWAM-α led with a 55.3% success rate.
Why it matters: The benchmark quantifies the capability gap between general large models and post-trained policies, helping readers judge which task lengths and manipulation precisions each approach suits.