PteroSim 0.5.0 has been released. The core update is that .glb props become physics objects: they can be dragged into the scene or generated from Python, and they fall, collide, and can be grasped by the gripper.
The author turned to high-quality Gaussian Splats of real sites to build augmentable, relightable photorealistic simulation environments, because policies trained in animated simulation were hard to transfer to real machines. The pipeline, refined over hundreds of iterations, converts iPhone scans into Splats plus meshes, selects robot-operable assets, and splits them by unit to generate detailed assets placed back at their original positions in the Splat. Physics and rendering use MuJoCo Warp, and the robot is a Franka Panda.
Apple's research team published the paper "Perceptive Humanoid Motion from Simulation" on arXiv, unveiling HALO, an autonomous control system for humanoid robots, and simultaneously open-sourcing the training and deployment code.
Jianzhi has launched the Dex 2.0 data glove and the DAS Ego 2.0 first-person head-mounted device, building a Human Data collection system covering first-person vision, near-field wrist vision, hand motion, and touch.
Professor Lan Xuguang's team at Xi'an Jiaotong University has proposed AffordSim, which connects open-vocabulary 3D affordance prediction to simulation trajectory generation. Starting from a natural language task description, it locates an object's functional regions, then combines cuRobo motion planning to generate executable manipulation demonstrations.
At CoRL 2026, NVIDIA launched the open-source world-action model series GR00T 2, comprising two versions, GR00T 2 and GR00T 2 Turbo. The former focuses on generalization and fine manipulation, while the latter has fewer parameters and can be deployed directly on the Unitree G1, with a single inference taking 124 milliseconds on the onboard Jetson Thor.
Why it matters: NVIDIA released the GR00T 2 series at CoRL 2026 and disclosed the scale of three types of pre-training data along with success rates for cross-embodiment transfer and few-shot adaptation, which can be used to assess the current capability boundaries of world-action models.
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.
Jianzhi Robotics announced that its DAS 2.0 series appeared on-site at the CoRL conference in Austin, USA, on November 9, including the DAS Dex 2.0 dexterous-hand data collection glove and the DAS Ego 2.0 first-person-view head-mounted multimodal collection device, which synchronously record visual, hand-motion, and tactile information through head-hand coordination.
On November 10, Vbot Vita Power released the Super Robot Dog DEV Geek Developer Edition, officially priced at 22,988 yuan, with a limited-time Double 11 promotional price of 20,988 yuan. The platform comes standard with a 360° lidar, HD stereo cameras, and a microphone array, supports Bluetooth, Wi-Fi, 4G connectivity and Type-C peripheral expansion, and offers six software foundations: spatial memory and navigation, following, local visual recognition, voice interaction, the RCP robot control platform, and an action package repository.
A team at the University of California San Diego proposed EgoThumb in an arXiv preprint, a wearable robotic thumb driven by a palm-side wide-angle camera and an imitation learning model that can autonomously grasp and stabilize objects, allowing users to complete tasks such as unscrewing bottle caps and stirring tea with one hand.
NVIDIA's GEAR team released its next-generation world action model GR00T 2 at the CoRL conference. It can simultaneously predict future frames and robot actions. The model, data, simulation environment, and collection hardware will be open-sourced progressively, and 100,000 hours of human data have already been released first.
Why it matters: NVIDIA GR00T 2 pretrains on a mix of human video, UMI, and teleoperation data, and reports quantitative success rates for cross-embodiment transfer, which can be used to assess the data strategy for robot foundation models.
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.
Ant Group's Lingbo has open-sourced LingBot-Touch, a tactile pretraining foundation model that uniformly encodes magnetic, visuotactile, and piezoresistive tactile signals into tactile tokens, which are then passed to the action model. Team experiments showed that when π0.5 was directly fed raw tactile readings on a zipper-pulling task, its success rate dropped from 39% to 15%; after connecting LingBot-Touch, it rose to 59%.
Why it matters: The original provides a controlled experiment showing that feeding raw tactile signals into the action model instead lowers the success rate, along with the improvement after unifying tactile tokens, which can serve as a reference for tactile representation design.
Point clouds are discrete geometric samples of three-dimensional space. Unlike regular images, they must be converted into a computable BEV representation through structured processing such as voxels or pillar bucketing. The article reviews point cloud algorithms along the chain from data fundamentals to models, engineering deployment, and quantization, and organizes common computation modules.
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.
Paris-based physical AI training data platform Physicl announced it has raised a $10 million seed round led by Inovia Capital, with participation from US Venture Partners.
A paper proposes that how tactile data is represented may matter more than resolution. The authors pretrained a tactile encoder, LingBot-Touch, on various sensors ranging from 3x3 glove arrays to 240x240 optical fingertips, and tested it with 4 strategies on 26 real tasks.
Tutor launches TRACE, which uses deployment-time compute scaling for robot learning: starting from a small number of Tutor demonstrations, a frontier large language model agent reconstructs the task in simulation and trains reliable, human-speed policies. TI-0.1 is the first model trained with TRACE. USV believes this further shows that deploying robots in the field accelerates learning, with each deployment generating training data, giving Tutor a new way to make robot learning continuously accumulate.
Ken Goldberg's team at UC Berkeley and Spencer Huang, NVIDIA's head of robotics engineering, proposed an "agentic robotics" method that lets AI agents generate, test, and correct robot control programs in simulation without additional data collection.
Why it matters: The report describes a method for using multi-agent systems to generate, test, and correct robot programs in simulation, and presents real-world data from warehouse sorting.
Embodied intelligence data collection centers on teleoperation architecture and the Demonstration data pipeline, explaining how to collect demonstration data through teleoperation. The content focuses on the practical data collection workflow, covering the setup of the teleoperation architecture and how the demonstration data pipeline is organized.
The article discusses why injecting only Gaussian white noise in ROS2 mobile robot simulation is insufficient for validating fusion algorithms, focusing on systematic errors such as bias, scale, and mounting misalignment that cannot be averaged out by filtering, and provides a complete approach for injecting systematic errors into odom, lidar, and IMU.
In the second half of 2026, as embodied intelligence funding cooled, Maxinsights, Jingshuo Technology, and Jianzhi Robotics completed a $30 million seed round, a several tens of millions of dollars Pre-A round, and a $100 million A+ round, respectively, within ten days. None of the three companies builds complete robots; their core teams have backgrounds in autonomous driving data closed loops, and they enter data supply through a global collection network, a data platform spun off from WeRide, and a multimodal collection production line, respectively. The article points out that data volume does not equal data value, and long-term competitiveness depends on data production and cross-embodiment reuse capabilities.
The authors release the FHNW Muttenz dataset, covering a contiguous 10 km street network and two mobile mapping campaigns about 1.5 years apart, with ground-truth poses accurate to 6-9 mm in position and orientation error below 0.1°. Using this dataset, they systematically evaluate sparse-feature visual localization; with daytime imagery and pre-calibrated cameras, median translation error is 1-5 cm and rotation error is 0.03-0.1°. The authors note that, provided the scene representation is precisely georeferenced, visual localization can approach survey-grade GNSS positioning accuracy and additionally provides full orientation information.
VOMMI is a portable demonstration collection and learning framework that connects portable RGB demonstrations to VLA post-training through offline trajectory reconstruction and online visual-motion conditioning. R2-VO refines trajectories using sparse geometric anchors, reducing the average absolute trajectory error of body and hand streams by 24.6%; a policy post-trained only on portable demonstrations has an 18.2% lower base velocity error than a policy trained on robot-collected demonstrations, with comparable end-effector translation accuracy.
The study constructs a geometric visual pretraining dataset containing only a plane, an object, and a cube hand, with trajectories generated automatically and scenes without textures or backgrounds, so that the policy is first exposed to the geometric relationship between the hand and the object. The authors evaluate ACT on three simulated robots with five manipulation tasks each and on three real-world robot tasks; in most combinations, fine-tuning from the geometric prior yields higher success rates in early training than training from scratch, and requires only a small number of task demonstrations. The results indicate that highly simplified geometric scenes can also provide an initialization that transfers across robots and to real-world tasks.
The authors propose LAG-Fusion, a latency-aware guidance fusion framework for asynchronous multimodal diffusion policy composition that allows modality-specific policies to run at their own inference rates and provide denoising guidance whenever output is available. To keep asynchronous composition consistent, the authors derive a reference-frame rebasing rule for diffusion variables under relative action representations, so that delayed guidance is aligned before fusion. In contact-rich manipulation, the authors compose a low-frequency vision policy with a high-frequency force policy; experiments with heterogeneous modality latency show better responsiveness and task performance than synchronous fusion and a purpose-designed force-aware baseline.
The authors propose DexTeleop-0, a bimanual dexterous teleoperation framework that uses fingertip tactile feedback to overlay robot-side contact corrections on human tracking commands, mitigating kinematic mapping errors and the inability of vision alone to regulate contact force. The method computes localized corrections in real time via contact-stiffness approximation and the kinematic Jacobian, and adds per-contact force tracking, soft regularization of predicted contact moments, and bounded hand-joint residuals.
The authors propose SGM-SLAM, a distributed SLAM framework for multi-robot collaboration in GPS-denied, communication-constrained environments, where robots can each initialize their coordinate frames independently and share only limited spatial overlap. The method uses fused RGB-LiDAR point clouds to build a scene graph containing a semantically segmented point cloud layer and a discrete bounded object layer, and performs collaborative matching by exchanging and matching object data with neighboring robots, using multi-step data exchange and optimization to improve communication efficiency. The authors validated effectiveness and communication efficiency on real-world datasets containing limited trajectory overlap, viewpoint variation, repeated structures, and degraded visual conditions.
CLAW is a fully end-to-end self-supervised framework that jointly learns a world model and continuous latent action representations directly from videos without action annotations. It uses adversarial latent regularization to prevent the latent variables from encoding future observations, so that they capture action structure, and a diffusion model models environment dynamics as a distribution over future observations rather than a single deterministic rollout. The authors say this latent action world model supports imitation learning from observation and goal-directed planning, and outperforms existing methods in multi-task, multi-embodiment experiments.
The authors propose XIT, a sampling-based planner that formulates active gas distribution mapping (GDM) as a next-best-trajectory informative path planning problem, growing a single shared tree toward multiple frontier goals while taking informative gas samples along the way.
The authors propose the Visual Learned Distance (VLD) framework, which decouples perception learning from policy learning for image-goal navigation: first, a distance-to-goal predictor is trained self-supervised on about 3,000 hours of internet and simulated video; then, in simulation, a reinforcement learning policy is trained using privileged geometric distances perturbed by GeoNoise; at deployment, VLD predictions are used zero-shot.
Augmentron is a geometry-aware data augmentation pipeline that generates reusable scene representations and renders them according to recorded camera poses, producing environment and workspace variations that are consistent across time and multiple views while preserving the robot, task objects, actions, and annotations.
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 team developed a wearable haptic input/output device that integrates a 36-channel electrotactile array with capacitive touch sensing, enabling stimulation locations and user-reported perceived locations to be aligned within the same body-referenced coordinate frame. Experiments with 24 participants showed notable individual differences in electrotactile and thermal percept maps, but both modalities retained electrode-pair-specific spatial structure across participants.
SURE is a data-efficient vision-language-action (VLA) training framework that models spatial uncertainty with a lightweight transposed-convolutional probabilistic action decoder and uses Soft-DTW alignment supervision to accommodate temporal variation across demonstrations. On LIBERO, RoboTwin 2.0, and real manipulation tasks, SURE matches or exceeds baseline VLA manipulation success rates using only 10% of demonstration data or 20% of training steps.
MDF++ is a visuo-tactile world action model that uses a diffusion backbone to learn latent representations of multisensory observations end-to-end, addressing the problem that existing world action models rely on frozen pretrained encoders whose representations are not well suited for control. The authors propose a multi-task learning strategy that leverages cross-modal correlations to learn task-relevant representations, avoiding the feature collapse caused by naive joint training. On contact-rich manipulation tasks including real-world bolt screwing and nut mating, it outperforms world action models built on frozen pretrained representations by up to 50%.
SUGARv2 proposes a unified framework for learning diverse visual humanoid loco-manipulation skills through interaction reference synthesis and a shared training pipeline. The method adapts reusable task templates to different spatial configurations to synthesize coordinated robot and object motion references at low cost, and uses reference-based imitation as a common learning objective across tasks, first training privileged teachers and then training visual policies through teacher distillation and reinforcement learning.
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.
The authors propose A2-MEAS, an embodied multi-agent framework that uses code as a scaling interface, combining corrective language, spatial evidence, execution history, and action proposals into correction codes; during training, MetaBrain provides supervision and runtime guidance.
Why it matters: The framework uses code as a scaling interface, injects corrective language and execution history into the policy, and provides quantitative comparisons across benchmarks and real-robot tasks.
The paper formalizes the attribution problem of inconsistencies between scans and models in LiDAR-inertial odometry and BIM fusion as fault identification over three fault classes, and derives a geometric criterion for determining whether attribution is feasible. On cases that are detectable but not attributable in a two-dimensional study, the method based on this criterion had a conditional error rate of 0.83%, compared with 31.67% for a baseline using uncertainty-based abstention; across 5,600 attributable cases, it produced 5,574 unique attributions with no erroneous attributions.