harpreet

a robot can recognize every object in a room and still not know how to open a single drawer knowing there's a cabinet is scene understanding. knowing which handle to grab, which direction to pull, and that it's a drawer not a door is affordance understanding almost no dataset has it in 3D SceneFun3D: 14,800+ functional elements across 710 laser-scanned indoor scenes. every handle, knob, button, and switch labeled with its affordance, motion axis, motion type, and a natural language task description each scene's 3D point cloud with its iPad video recordings in fiftyone. the affordance annotations live in 3D and project into the video frames check it out here huggingface.co/datasets/Voxe…

Help @Voxel51 make it to the top 10!!

84% of these robot rollouts failed. all 4,000 are in the dataset, labeled. 632 successes and 3,368 failures across 28 manipulation tasks and 99 scenes. every take has a wrist camera, two exterior cameras, and depth. i parsed RoboLab-EgoX into FiftyOne so you can filter to the failures and watch them from every angle. huggingface.co/datasets/Voxe… #PhysicalAI #Robotics

four depth cameras in a ring. the paper asks whether more of them help. the multi-RGBD inertial set uses 4 Intel RealSense D455s, a MicroStrain IMU, and an Ouster-64, with hardware sync under a millisecond. 6 sequences: glare, low light, textureless walls, dynamic occlusion. RA-L 2025, ICRA 2026, with Hyundai Motor Company. reference trajectories are LiDAR-IMU SLAM in TUM format. github.com/seungsang07/multi… #PhysicalAI #Robotics

most robot training runs are wasted on episodes that teach the policy nothing. the fix isn't a bigger model. it's finding the failures that matter and training on those. that's what we built. it's live on product hunt today and we need your help climbing 👇 producthunt.com/products/vox…

a person, a cobot, and a mobile robot can reach the same cell together. the ULTIMATE workshop twin is a synthetic industrial workspace with 3 synced RGB-D streams, a cobot, an AMR, and motion-capture human motion. it has 4 hazard events in 12 ROS 2 Humble bags, 20 GB. zenodo.org/records/17235460 #PhysicalAI #Robotics

falls and daily actions, frame by frame, including people who use wheelchairs. SAFER-Activities is 66 hours of multi-camera monitoring. 85K frame-level action instances across 30 classes, with a wheelchair-user subset. ECCV 2026. tied to healthcare monitoring. on Hugging Face. arxiv.org/abs/2609.08038 #PhysicalAI #MedicalAI

point a lidar at a mirror and it measures a room that isn't there. in an elevator, the cabin appears to extend into solid wall. MirrorSentinel runs a ZED2 camera and an Ouster lidar through 7 cabins. each cabin was measured by hand, so you can check the real boundary against what the sensor reports. i parsed it into FiftyOne so you can scrub the camera and the point cloud together. huggingface.co/datasets/Voxe… #PhysicalAI #Robotics

follow one object in a point cloud without a class name. TETrack3D tracks a single 3D object with temporal key-value caches and a state that evolves over time. github.com/NJU-PCALab/TETrac… #PhysicalAI #ComputerVision

can't be bothered typing slack messages so now i'm gonna just send voice memos

swap the clothes, the person, or the room while the video keeps playing. Vidu-S2 does real-time interactive editable spatial video generation at 720p, with dynamic references and replacement of style, clothing, character, and background. arxiv.org/abs/2609.11638 #PhysicalAI #ComputerVision

similarity metrics trained on photos miss ultrasound speckle. UltraPIPS is a library of perceptual metrics for B-mode ultrasound. standard LPIPS backbones trained on natural images miss speckle and other acoustic statistics. arxiv.org/abs/2608.26033 #PhysicalAI #MedicalAI

the gantry already knows where it is. the radios have to catch up. RoboLoc-G is localization and tracking on an industrial 3-axis CNC gantry with sub-centimeter measured position. it includes 60 GHz TI mmWave radar (IWR6843ISK and IWR6843AOP), Wi-Fi FTM ranging, Pozyx UWB with 9-axis IMUs, a 10-camera OptiTrack system, and 2 external video cameras. ROS 2 bags cover 8 trajectory patterns. zenodo.org/records/15989282 #PhysicalAI #Robotics

most benchmarks in this domain test against synthetic assumptions. real-world deployment breaks on edge cases nobody captures. RGB-P Wild captures the ground truth: 10,000 paired rgb and 4-angle linear polarization sets (0°, 45°, 90°, 135°) for autonomous driving perception under glare, transparency, and specular reflections. arxiv.org/abs/2610.25396v1

a wrong 3D label might be the detector, or it might be the lift. Semantic Lifting for 3DGS aggregates multi-view 2D detector masks into 3D Gaussians with frustum visibility weights and spatial hysteresis, so lift error and detector error stay separate. github.com/ivanver02/semanti… #PhysicalAI #ComputerVision

one short autonomous-racing bag: color, depth, and a fast IMU. Race16 is a single ROS 2 MCAP, about 339 seconds, extracted to 20K rows (2.67 GB). RGB at 848 by 848, 16-bit depth near 30 fps, IMU near 200 Hz, odometry, position commands, and SO3 commands. no labels. huggingface.co/datasets/dmod… #PhysicalAI #Robotics

a world model can fake memory by barely moving. R2M-Bench checks revisit consistency in video world models with MemoryGain and NMR. it compares the first visit with the return and flags a slow-motion shortcut. arxiv.org/abs/2608.27328 #PhysicalAI #ComputerVision

language features and visual features vote. entropy picks the winner. UCF-Net fuses CLIP language-aligned semantics and DINO visual-structure priors with entropy gating for deepfake detection. it outputs a class and a confidence score. arxiv.org/abs/2609.07670 #PhysicalAI #ComputerVision

one moving camera, and nobody labeled the depth. Depth Anything V4 reconstructs a dynamic 4D scene from monocular video with Riemannian Flow Matching on 4D Gaussian Splatting. no human-annotated depth labels. it outperforms Depth Anything V3 and per-scene 4D-GS. arxiv.org/abs/2608.18388 #PhysicalAI #ComputerVision

tilt a lidar 45 degrees and every SLAM assumption tilts with it camera-lidar overlap shrinks, upper beams go sparse, standard odometry drifts. that's exactly how compact mobile mapping rigs are built in the real world. the lidar tilts so it sweeps more vertical structure. but almost no benchmark tests this configuration YUTO MMS from York University: a tilted 32-beam lidar, a 6-lens panoramic camera, and GPS/INS with cm-level ground truth driven through Toronto. every lidar point is RGB-colorized from the nearest panoramic frame, not a synthetic colormap loaded as mcap in fiftyone for your viewing pleasure scrub the timeline and watch the world-frame 3D map build itself progressively alongside the panoramic camera, GPS track, and IMU telemetry checkout the dataset here: huggingface.co/datasets/Voxe… it's running as a live space too, nothing to install: huggingface.co/spaces/harpre… #mcap #skills #computervision #ai #artificialintelligence #machinevision #machinelearning #physicalai