Search - metacurate.io

Train and Deploy Action Recognition Models with Roboflow

Roboflow has added tools to train and deploy models that identify actions in video and mark when each action starts and ends. Tests on three public datasets found higher accuracy for V-JEPA 2.1, while Cosmos 3 Edge sometimes placed action boundaries more precisely.

8 mins read Listed 3 days ago in RSS: Roboflow Blog

github.com Tools and repositories

GitHub - starman-tech/mousepods: Your AirPods are a mouse: AirPods 3 head tracking on Linux, and the first public documentation of Apple RemoteHID (HID over AACP) · GitHub

A Linux-focused project documents how AirPods 3 expose motion sensors via Apple RemoteHID, enabling head-tracking and a mouse-like cursor through a local app. It details hardware protocols, sensor data formats, calibration, and multiple operation modes, with tests and notes on limitations.

5 mins read Listed 9 days ago in reddit/ClaudeAI/used claude code to figure out how airpods 3 send head tracking data, turns out nobody had done it before

docs.menlo.ai Documentation

Mechanical · Menlo Docs

The document details the Asimov 1 mechanical design, including its modular head-torso-pelvis structure, limb DOFs, materials, rated and peak loads, joint axes and limits, coordinate conventions, and actuator specifications. It notes upcoming interactive visualization and updates to reflect current design.

8 mins read Listed 24 days ago in reddit/singularity/Building a humanoid for agentic robotics

huggingface.co Tools and repositories

nvidia/PhysicalAI-Human-1M-100K · Datasets at Hugging Face

Human-1M is a large egocentric dataset of human work, combining video, language annotations, and head and hand motion for physical AI. The page describes its contents, collection scale, quality checks, data format, and CC BY 4.0 license.

3 mins read Listed 2 days ago in X

arxiv.org Scientific research

[2610.34186] HuGe: Scaling Humanoid Generalist Vision-Language-Action Models with Egocentric Experience

HuGe presents a family of humanoid vision-language-action models pre-trained on egocentric data, using a three-stage pipeline to transfer policies to humanoid robots and achieve strong simulation and real-world results, including improved pick-and-drop performance on a Unitree G1.

1 min read Listed 21 days ago in Hugging Face Daily Papers

arxiv.org Tutorial

[2610.25288] Vision as Proprioception: Estimating Humanoid Robot Pose from Video

HumanoidPose treats vision as proprioception to recover full-body robot joints from video, by merging human-pose features with robot kinematics, retargeting to sixteen humanoids, and using self-training on unlabeled videos to improve a 2D robot-pose detector; it achieves notable accuracy gains and better generalization across real platforms and unseen embodiments.

1 min read Listed 28 days ago in Hugging Face Daily Papers

arxiv.org Scientific research

[2610.42096] CausalWAM: Learning the Causes of Actions in World Action Models

The paper introduces CWAM, a method to separate action-causing latent variables from predictors in world action models, proving identifiability under sparsity and showing improved efficiency in closed-loop tasks through experiments on synthetic and real-world robots.

2 mins read Listed 16 days ago in Hugging Face Daily Papers

arxiv.org Scientific research

[2610.31716] Does LeWorldModel Preserve Local Action Geometry?

The paper measures how closely LeWorldModel replicates small action changes using Gram matrices across three perturbation scales and two environments, finding scale-dependent action sensitivity and that larger perturbations reduce discrepancy; embedding geometry varies with scale, and the impact on control remains unresolved.

1 min read Listed 24 days ago in Hugging Face Daily Papers

leggedrobotics.github.io Tutorial

Humanoid Mobility via Attention-based Contacts

A dual-network system operates at 50 Hz to plan and execute humanoid motion. The contact planner forecasts 9 body-part contacts over 1.5 seconds using height maps, proprioception, and a goal; the whole-body controller translates the 9×3 plan into 23-joint actions without reference motion at deployment.

1 min read Listed 20 days ago in reddit/robotics/Humanoid Mobility via Attention-based Contacts (ETH Zurich)

link.springer.com Documentation

Vision-based touchless gesture recognition for clinical HCI: deployment challenges, trade-offs, and future directions | Artificial Intelligence Review | Springer Nature Link

Supplementary tables quantify gesture-recognition accuracy gaps between lab tests and clinical use, and compare processing speeds to a 20 FPS real-time threshold, citing Wang 2025 and Jacob 2013 for the benchmark.

1 min read Listed 18 days ago in RSS: Latest Results