CyberRobo ✓

[video] No wonder they say teleoperation is dead Peking University’s DAGroup released HumanNet, a massive 1 million hour dataset of human-centric videos that turns everyday internet footage into gold for embodied AI. It’s got first-person and third-person views, super detailed

[video] Scaling Laws for Humanoid Robots？ A cool new CVPR 2026 project called Humanoid-GPT uses a GPT-style AI model trained on 2 billion motion frames. It lets humanoid robots copy new movements instantly (zero-shot) in real time,smoothly and stably. >Handles both fast dynamic

[video] NVIDIA just open-sourced GR00T-VisualSim2Real A powerful visual sim-to-real framework for humanoid loco-manipulation. Train entirely in simulation with VIRAL & DoorMan, then deploy zero-shot on real Unitree G1 using only RGB cameras + proprioception. Master tasks like walking,

[video] Ah… human hand is already a perfect natural tactile glove Add just one low-cost wristband: it reads forearm sEMG to capture real contact forces and fine fingertip forces from human natural hand movements. (Muscles don’t lie.) No sensors, non-intrusive… This turns our innate

[video] Perhaps the "GPT-3 moment" for humanoid robots lies within vast amounts of human video data. UC Berkeley (Abbeel + Malik group) just released Do as I Do: a pipeline that takes monocular RGB human videos (ego + exo)->reconstructs hand-object interactions->retargets them to

[video] NVIDIA MotionBricks brings humanoid robots closer to natural, fluid movement The system combines a modular latent generative model with Smart Primitives, packing more than 350,000 motion skills into a single backbone. It runs in real time at up to 15,000 FPS with just 2ms

[video] Another major Chinese company had open-sourced a humanoid robot dataset-- EgoLive from JD. This is a large-scale first-person perspective dataset recording real-world human tasks. It contains 1680 hours of high-fidelity binocular video data at 60fps, involving 65,866 data

[video] Amazing…The Amazon FAR team just got a Unitree G1 humanoid autonomously climbing ladders. LadderMan, it can stably climb ladders of varying rung spacing and angles, while performing real tasks like adjusting wall art, handing over boxes, or screwing in bulbs--all with keyboard

[video] We are entering the "oil refining" era of embodied AI This data engine, named EgoInfinity, distills massive video datasets,such as YouTube and Action100M (comprising over 100 million videos)--into structured 4D hand-object interaction data used to train humanoid robots.

[video] Dream2Flow from Stanford University marks a major breakthrough: Simply put, humanoid robots can learn physical movements from videos generated via text/images, and subsequently execute tasks successfully in real-world environments. This significantly reduces the effort required

[video] NVIDIA just dropped RoboTTT--a way to give robots much longer “memory” of what’s happening, up to about 5 minutes of continuous action (way beyond the usual single-step or short-history models). The results are pretty cool: >Watch one human video → robot can imitate it right

[video] HumanEgo is a really cool project They can train a working robot policy using just ~30 minutes of human first-person video (even 15 minutes can work decently). 92.5% success rate on real robots, and zero robot data needed. What I like most is they showed human videos actually

[video] Just discovered SimToolReal:an insane breakthrough in robot tool manipulation No more task-specific training,it uses sim-to-real RL to handle unseen tools zero-shot. Hammers, scissors, you name it,over 120 real-world demos crushing it. more: Mind-blown

[image] Cool tactile sensor breakthrough from Queen Mary University of London. They created a soft material that changes color under pressure, letting a regular camera capture real-time, high-res maps of touch, strain, and force--no heavy algorithms or dense sensor arrays needed. For

[video] How do you teach a humanoid robot to walk, see, and manipulate objects using only language commands--without tons of expensive real-world data? VLK's answer: Scan real rooms with an iPhone, rebuild them in 3D, synthesize 48k vision-language-motion trajectories, and train a

[video] A humanoid butler that works while moving Researchers from NTU, PKU, BAAI and HKUST(GZ) have released ω-0(OMEGA-0), a latent predictive world-action model for concurrent humanoid loco-manipulation. Instead of splitting walking and manipulation into two systems, it directly

[image] Humanoid robots are finally getting “real muscle memory”💪 A Seoul National University team just created an intelligent artificial muscle that can move AND sense at the same time--exactly like your own biceps. They linked two special rubber-like materials in series (think

[video] Traditional teleop data is expensive and hard to scale. NVIDIA & UT Austin’s HumanoidMimicGen changes that-- 1 demo generates thousands of diverse full-body trajectories, hitting 89% success and lifting real-world performance from 51% to 71%.

[video] Humanoid robots often look capable, but break down on contact. This project 👉 adds a key piece: predicting future touch during action — Touch Dreaming. Instead of see → act it becomes see + touch → predict contact → act That shift shows up in tasks

[video] RLDX-1 is a new robot hand foundation model recently released by the Korean company RLWRLD, with a strong Dexterity-First focus. It was designed from the ground up by targeting real-world tasks where robots commonly fail,such as pouring coffee, grasping objects, and manipulating

[video] NVIDIA Cosmos 3 is an omnimodal world model: within a unified architecture, it can understand and generate language, images, video, audio, and actions. Simply put, it enables humanoid robots to truly understand physical laws --knowing when things might slip or break, and how to

[video] Humanoid robots can finally think before they act AGIBOT’s τ0-WM unifies video prediction + action generation in one model. It mentally simulates options, evaluates risks, and picks the best move,just like humans. Huge leap for precision robotics

[video] Cool Universal Manipulation Exoskeleton What impresses me most about UME-Exo (Ant Group + Stanford) is its real-time transparent torque haptic feedback--you can directly “feel” the resistance the robot arm encounters, just like your own hand. This boosts embodied AI in

[video] Many people know that dexterous robot hands are a bridge to robot intelligence… But for most uncertain and complex household tasks, hands alone are far from enough. What’s truly needed is a closed-loop capability that integrates high-level semantic decision-making, Whole-Body

[video] A wave of affordable wheeled dual-arm robots has recently emerged... It’s really impressive that, for just $1,000, you can now assemble a robot capable of autonomous sorting and transport for lightweight, simple tasks.

[video] An interesting grasping project called HUG from NYU the hand is the bridge and entrance of Physics AI--so,their idea is the best teacher for robot grasping is human demonstrations. Using smart glasses, they collected ~28 hours of first-person egocentric videos to build the

[video] Another Chinese big company just open-sourced a robot foundation model. Xiaomi tech dropped Xiaomi-Robotics-1:ready-to-use robot open-source foundation model It’s trained on 100,000 hours of embodiment-free (UMI) trajectories spanning more than 1,700 scenarios( household,

[video] Tencent released an open-source project: Hy-Embodied-0.5-VLA--an engineering-driven embodied AI system focused on real-world robot learning With 10k+ hours of high-precision egocentric UMI data, strong innovations in cross-embodiment transfer and offline RL (FlowPRO), it goes

[video] Xiaomi Robotics has introduced Xiaomi-Robotics-U0, a 38B-parameter autoregressive world foundation model. Generative models like this are increasingly becoming “data engines” for robots. U0 brings text-to-image generation, image editing, multi-view embodied scene generation,

[video] Controlling a robot with Claude Agent? Waddle Labs’ approach stands out from the usual data-hungry VLA/WAM route. Their LLM agents take a prompt, watch the camera, and write editable control code--great data efficiency for new tasks or hardware, reusable skills, multi-agent

[video] A rather interesting experiment: Handroid Same set of 27-DoF modules can act as either an anthropomorphic dexterous hand or a desktop humanoid. Finger joints literally become the legs. 0.33 m tall, just over 2 kg, fully 3D-printed. It does grasping, cup stacking, in-hand cube

[image] How do socially interactive humanoid robots alter human hormones? A Drexel study had people interact with two versions of Pepper. One made eye contact, nodded, and gestured. The other spoke the same lines with almost no body language. The expressive robot built trust faster.

[video] HiFi-UMI-2K: Human data accurate enough for real robots Beijing Simple AI open-sourced 2,000 hours of high-fidelity bimanual manipulation data. >Built at 20,000+ hours. >4.32M+processed episodes >480+collection scenes >98%+valid trajectories The surprising result: policies