Fly across one island that holds my field work: mountain trails, caves and mines, a volcano, a disaster site and a home. 33 stops, each with the real footage.
Enter the expedition →Part of the Gemini Robotics team, building one model that sees, reasons and acts on any robot body. I focus on its intelligent whole-body control capability.
About this project →Gemini 4 Argon can build its own simulations, practise in them, and then control real robots. I worked on these two launch demos:
Argon watched a single video of a person playing cornhole, wrote a physics simulation of the robot and the board, trained the throw there, and the real Apollo sank 7 throws in a row.
Argon built a MuJoCo twin of a foot-pedal trash bin, practised there, then had Spot balance on three legs, step on the real pedal and drop a soda can into the open bin.
Six years at ETH Zurich’s Robotic Systems Lab, teaching the ANYmal robot to see, walk and explore: work that appeared in Science Robotics four times.
Reinforcement learning that lets legged robots see the ground and still trust their legs.
Learning from recorded motion so robots move naturally, then adapting it to rough terrain.
elevation_mapping_cupy: fast, GPU-based terrain maps for legged robots, and planning on top of them.
Legs that also handle things: opening doors, balancing an arm, catching a ball.
First place and the $2,000,000 prize in the 2021 Final.
Taking legged robots to volcanoes, planetary analog sites and city streets.
Every day a bot picks a handful of new robotics and AI papers and writes a plain-language summary of each, in English and Japanese. These are the latest.
DAMP: Humanoid Locomotion via Denoised Belief Learning and Adversarial Motion PriorsThis paper presents DAMP, a reinforcement learning framework for robust and naturalistic humanoid locomotion over challenging terrains without exteroceptive perception. By leveraging RNNs to infer privileged latent states and combining them with adversarial motion priors, it achieves stable and goal-directed blind traversal across complex environments.
Being-M0.7: A Latent World-Action Model for Humanoid RobotsBeing-M0.7 is a latent world-action model designed for humanoid loco-manipulation. By pre-training on over 10,000 hours of human video and motion data and transferring learned visual-motion priors via mid- and post-training, it enables coordinated whole-body humanoid control across diverse manipulation tasks.
VioLA: Learning Generalist Humanoid Control Policies from Human DataVioLA introduces a generalist humanoid policy learned from large-scale human demonstrations. Instead of predicting low-level joint actions directly, it predicts whole-body target trajectories tracked by lower-level controllers, enabling out-of-the-box instruction following without task-specific teleoperated demonstrations.
Teaching a Robot Dog New Tricks: Diverse Quadruped Skills via Combined Reinforcement and Imitation Learning with Adversarial Task SelectionThis paper presents a three-stage framework combining RL and imitation learning with adversarial task selection to train a single quadruped policy capable of diverse skills like walking, digging, and hopping. It mitigates multi-task gradient conflict by distilling multiple teacher policies into a unified student policy.
A Balanced Data Diet: Addressing the Exploration Bottleneck in Mega-Scale RL for Robot ControlThis work tackles the exploration bottleneck in mega-scale parallel RL for agile locomotion and dexterous manipulation. By showing that uniform reset distributions hinder training efficiency, it proposes an adaptive reset sampling strategy that significantly improves learning progress and policy performance.
OmniDex: Scaling Dexterous Hand Grasping to Diverse Cluttered ScenesOmniDex introduces a massive simulation dataset and benchmark for dexterous grasping in cluttered environments. Using a scalable seed-and-filter strategy, it generates over 2.6 million scenes and 0.4 billion grasp ground truths, providing rich training data to overcome real-world data scarcity.