Humanoid robot research and AI models
HumanoidBeat coverage of humanoid robot research: control methods, robot foundation models and new training approaches, as described by their authors. Newest first; each story links to its sources.
- Researchers turned internet videos into 5 million moves a Unitree G1 can actually perform, and released the model
Zhejiang University, Unitree and partners have released Movens-5M, 5 million robot-executable motion sequences built from online video, and Movens-Zero, a model that drives a Unitree G1 from text, speech, video or music. - Apple researchers train a Unitree G1 to move through unseen rooms using mostly simulated data
Apple's HALO system ran entirely on the robot and completed 91% of 360 single-task trials in six real environments, with no teleoperation data. Chained tasks were harder, averaging 70%. - Open source humanoid motion trackers now train in minutes on one GPU after Berkeley's SuperMimic
A UC Berkeley paper and a string of community projects built on it cut training for a Unitree G1 whole-body motion tracker from days on GPU clusters to a few minutes in simulation. The community results have not been tested on real robots. - Boston Dynamics puts Atlas on wheels to push its motors to the limit
Boston Dynamics fitted scooter wheels to Atlas's ankles and retrained its whole body controller to stress test actuators. It says the wheeled version is not a product. - Nvidia releases Isaac Lab 3.0 and GR00T 2 at CoRL as robot training shifts to dexterity
At the Conference on Robot Learning, Nvidia launched Isaac Lab 3.0, the GR00T 2 world action model and new benchmarks for precise manipulation tasks. - What is a vision-language-action model? How robots turn words into movement
A VLA model takes camera images and an instruction and outputs robot actions. RT-2, Gemini Robotics, pi0 and Figure's Helix are the best-known examples. - KIST KAPEX humanoid heads toward LG commercialization after science award
KIST’s KAPEX humanoid, built with more than 90% domestically made core parts and able to walk while carrying over 20 kg, has its technology path into LG after Lee Jong-won’s November science award. - RoboColosseum's Opening Results: No AI Policy Completed a Unitree G1 Humanoid Task
In RoboColosseum's real-world tests, no evaluated AI policy completed any of five Unitree G1 humanoid tasks, with average progress of 13.3% to 23.2%, its paper says. - LiberAI's Liber-0 Completes 36% of Unseen Robot Tasks Zero-Shot
LiberAI says its Liber-0 world action model, pretrained on 2.2 million hours of human experience, succeeded in 36% of trials on 60 tasks it was not trained for. - South Korea Opens First AI Humanoid Data Factory at KIMM to Train Robots for Real-World Tasks
South Korea has opened its first dedicated facility for producing training data for bipedal humanoid robots, placing hands-on robot learning at the center... - Researchers teach a Unitree G1 with a 20-joint hand to catch objects it never saw in training
A UC San Diego team trained a seated Unitree G1 in simulation to catch thrown objects using one head camera, succeeding in over a third of real tosses. - X-Era releases E0.5, an open robot model that predicts the 3D scene ahead, tested on a Zerith H1 humanoid
X-Era AI's E0.5 builds a robot control model on a 3D foundation model and trains it to predict future scene geometry. The team has released code and a checkpoint, and reports real-robot gains on a dual-arm system and a Zerith H1 humanoid. - Google opens Gemini Robotics ER 2, a high-level 'robot brain' model, to cloud customers in public preview
Google says its Gemini Robotics Embodied Reasoning 2 model, which plans and checks tasks for robots from arms to humanoids, is now in public preview on its Gemini platform for enterprise developers. - Why Reward AI's OM-1 humanoid parcel-delivery and stairs demo is getting attention
Reward AI posted video of humanoids running its OM-1 model to carry parcels from a vehicle and handle objects on stairs; the post passed 300,000 views on X within a day. - ETH Zürich's ACP policy lets a Unitree G1 climb ladders, vault and get back up after falls
Researchers at ETH Zürich's Robotic Systems Lab report a single learned policy, Attention-based Contact Planning, that plans where a humanoid's hands, knees, feet and pelvis should touch the terrain, letting a Unitree G1 climb, vault and recover from falls. - Amazon-led team teaches a humanoid to hold boxes against its torso and under its arm
A preprint from Amazon, UC San Diego and the University of Michigan describes Whole-Body Dexterity, a controller that lets a teleoperated Unitree G1 brace objects against its arms and torso, completing 45 of 50 real-world trials. - Figure says Figure 04 is in design lock and teases Helix 3.0 in a public all-hands
In a published all-hands recording, Figure CEO Brett Adcock said Figure 04 has reached design lock, its Index app has 37.5 million human videos, and previewed Helix 3.0. - Alibaba's Amap team releases λ0, a humanoid model trained largely on human video
Researchers at Alibaba's Amap and Token Hub released λ0, a whole-body humanoid model trained on a 500-hour human video dataset; they report 62.5% success on four real-world tasks with a Unitree G1. - XPENG shows IRON switching between handshakes, fist bumps and high fives
A new XPENG video shows its IRON humanoid adapting as a tester changes gestures mid-movement; XPENG credits its VLA 2.0 model, which it says re-evaluates the scene every 0.2 seconds. - Why UCLA's PhoneBot, a small humanoid run by a smartphone, is getting attention
UCLA's open-source PhoneBot uses an Android phone for sensing and computing in a 48 cm biped that its researchers say costs about $400 without the phone. - Unitree says it is building "Rehearsal", a self-improving loop for robot software
Unitree's Wang Xingxing said on 22 Oct 2026 the firm is developing "Rehearsal", a loop for robots to write code, test it in simulation and improve (SCMP). - Berkeley's SuperMimic trains a humanoid motion controller in three hours on one GPU, authors report
UC Berkeley researchers report that SuperMimic matched a baseline humanoid motion tracker that trained for seven days on 128 GPUs, using three hours on one GPU.
Other topics
Unitree · Figure · Agility Robotics · NVIDIA · China · South Korea · Europe · Pilots and deployments · Funding and policy