Peepo the peepo

Axis Weekly Last week, we made progress across the full robotics data loop, including task generation, simulation infrastructure, model training, and failure recovery. Key updates: - Task generation: We improved TaskGen with better automatic checker generation, stronger multi-embodiment support, and more efficient domain randomization to scale task diversity with less manual design effort. - Simulation infra: We continued improving MuJoCo verify/replay and scene-variant workflows, including fixes across data collection, multi-asset scenes, repeated loading/downloads, initial states, teleoperation, IK, and gripper control. - Model training: We confirmed that the new randomized tasks are learnable with sufficient data. In our current experiment, 500 demos successfully produced an executable policy, while 100 demos were not enough. - Failure recovery: We began building a recover-from-failure pipeline to collect and categorize gripper failure and near-failure states during grasping, which will later support more robust recovery policy learning. A closer look at this week’s progress🧵

Axis Weekly Last week, we made progress across the full robotics data loop, including task generation, simulation infrastructure, model training, and failure recovery. Key updates: - Task generation: We improved TaskGen with better automatic checker generation, stronger multi-embodiment support, and more efficient domain randomization to scale task diversity with less manual design effort. - Simulation infra: We continued improving MuJoCo verify/replay and scene-variant workflows, including fixes across data collection, multi-asset scenes, repeated loading/downloads, initial states, teleoperation, IK, and gripper control. - Model training: We confirmed that the new randomized tasks are learnable with sufficient data. In our current experiment, 500 demos successfully produced an executable policy, while 100 demos were not enough. - Failure recovery: We began building a recover-from-failure pipeline to collect and categorize gripper failure and near-failure states during grasping, which will later support more robust recovery policy learning. A closer look at this week’s progress🧵

Glad to see Axis inspiring the next generation to build in the physical world. Great things start small.

Noticed that object descriptions and reference images sometimes don't perfectly match the actual assets spawned in a task? This is because we’ve rolled out in-task randomization to increase data diversity and improve model generalization. The actual assets may vary, but the task goal always remains the same. Please focus on the task goal rather than the specific assets. Rich scenarios, diverse combinations of atomic skills, and extensive in-task randomization build the diversity of our data.

Which robotics project on Base do you like best?

Axis Weekly Last week, we made progress across the full robotics data loop, including task generation, simulation infrastructure, model training, and failure recovery. Key updates: - Task generation: We improved TaskGen with better automatic checker generation, stronger multi-embodiment support, and more efficient domain randomization to scale task diversity with less manual design effort. - Simulation infra: We continued improving MuJoCo verify/replay and scene-variant workflows, including fixes across data collection, multi-asset scenes, repeated loading/downloads, initial states, teleoperation, IK, and gripper control. - Model training: We confirmed that the new randomized tasks are learnable with sufficient data. In our current experiment, 500 demos successfully produced an executable policy, while 100 demos were not enough. - Failure recovery: We began building a recover-from-failure pipeline to collect and categorize gripper failure and near-failure states during grasping, which will later support more robust recovery policy learning. A closer look at this week’s progress🧵

Which robotics project on Base do you like best?