Google DeepMind presents two research systems that make robots more dexterous in the real world. The first, ALOHA Unleashed, shows that a robot can learn two handed tasks such as tying shoelaces, hanging a shirt, inserting a gear, and simple repair, by learning from human demonstrations and then acting on its own. It uses a learning recipe that mixes large scale teleoperation data with a modern generative model to predict the next motions of the hands.
Imitation learning is the method behind ALOHA Unleashed. In this approach, a robot learns by watching and copying examples given by people. The model uses a diffusion policy, a generative process that removes noise step by step to predict an action sequence, similar to how image tools create pictures from noise. This helps the robot plan sequences of precise hand motions for long tasks with cloth, laces, and small parts.
The second system, DemoStart, targets robots with many fingers and joints. It uses reinforcement learning, where the robot improves by trial and error with a simple success signal, mostly inside a physics simulator. The method builds an automatic curriculum, starting from easy states and moving to harder ones, then transfers the learned skill from simulation to a real robot hand with standard sim to real tools.
Together, these results point to a practical path for everyday dexterity. ALOHA Unleashed shows that scaling up demonstrations and using diffusion based policies can unlock two handed skills that were hard for robots, while DemoStart shows that a small set of demonstrations in simulation, plus trial and error, can produce skills that work on real hardware.
Limits remain. The ALOHA Unleashed study notes that performance drops for states that were not present in the training data, for example when the shoe flips or the laces tangle. This means reliability and recovery still need work before wide use in homes, hospitals, or factories.
Google DeepMind blog – Our latest advances in robot dexterity – 2024
Official announcement of ALOHA Unleashed and DemoStart, with examples such as tying shoelaces and hanging a shirt for ALOHA Unleashed, and simulation first learning and sim to real transfer for DemoStart.
arXiv – ALOHA Unleashed, A simple recipe for robot dexterity – 2024
Research paper showing that large scale teleoperation data, combined with a diffusion policy and a transformer model, enables a bimanual robot to perform long horizon tasks with deformable objects, including tying shoelaces and hanging shirts. The paper also discusses robustness and where the policy fails outside the training distribution.
arXiv – DemoStart, Demonstration led auto curriculum applied to sim to real with multi fingered robots – 2024
Research paper that learns dexterous skills in simulation from a few demonstrations and sparse rewards, builds an automatic curriculum, then distills the policy to a vision based controller that transfers to a real robotic hand.
0 Comments