VINO replaces a hierarchical point backbone with a Volume Transformer and redesigns its training objectives to learn semantic 3D features from unlabeled point clouds.
ABot-Recon reconstructs long videos incrementally with a 12-frame context, reporting 24.45 FPS at 6.71 GiB on KITTI-02 plus optional loop closure.
A first-person assembly sequence is synchronized to 6DoF camera poses and LiDAR depth. VedaAI frames it as lower-cost embodied-data capture on consumer hardware.
EMG-1 provides 307 hours of real-world work plus 10 hours of eight-channel EMG with optical motion capture; the authors report 25% lower fingertip error, rising to 35% when vision loses track.
RobotWorld tests multimodal agents on 84 simulation tasks, from manipulation to flight, with interaction budgets and hidden success checks.
Lightwheel contributes agentic QA and schema feedback to NVIDIA SimReady Central, validating contact, friction, mass, joints and articulated motion across more than 3,500 OpenUSD assets.
NVIDIA and GEAR Lab introduce VFlash, letting embodied models reason through visual plans and decode boxes, points and trajectories in parallel.
VFlash converts sequential visual-plan decoding for boxes, points and trajectories into a parallel process, giving physical-AI VLM models faster reactive capabilities.
Apple HALO runs on a backpack Mac mini and lets a Unitree G1 navigate unseen rooms, reporting 91% success across 360 real-world trials and six skills.
MIT LISLab introduces KeyFlow, where a coding agent improves a retrieval-based policy through simulated play so robots adapt from a few demonstrations without retraining.
NVIDIA VFlash decodes fixed-length geometry for object localization, robot trajectories, and dense 3D motion in parallel, claiming up to 12x faster action decoding, 5.1x end-to-end robot inference speedup, and 61% lower 3D trajectory error on edge GPUs.
The new Perceptron API converts first-person human and robot video into sub-second subtask timelines, per-hand activity labels, and 2D/3D hand tracks, with claimed Astra-level quality at about one-eighth the cost.
As an Amazon Associate, we earn from qualifying purchases.
