I used to think that once a robot learned to pick something up, you could just transfer the same behavior to another machine. After watching the Axis and Booster experiments, I realized how oversimplified that idea was. Franka can grasp a headphone case, but when switched to Booster, those exact movements likely won’t work at all. The two robots differ in arm length, number of joints, camera angles, and gripper design. Booster even has a dual-arm structure, with seven degrees of freedom per arm. The model must not only adapt to new visual inputs but also learn how to control this entirely new body. Previously, engineers had to stand beside physical robots, manually controlling them repeatedly to collect data. With limited machines, changing the environment or objects meant starting the data collection all over again. @axisrobotics solved this by first recreating Booster’s real workspace in simulation, then making the task accessible via browser so users worldwide could help operate it. Object positions, lighting, viewpoints, and backgrounds varied continuously—ultimately generating 42,046 simulated trajectories used to train Booster with a foundational policy better suited to its own body structure. The headphone case test results are clear: With only 10 real-world demonstrations, the robot failed to touch the target in all 20 consecutive attempts. After adding 50 corresponding simulated trajectories, it successfully touched the case in 17 out of 20 trials, grasped it successfully 6 times, and lifted it completely twice. In another test, Booster’s policy pre-trained on this data achieved better performance using just 30 real-world demonstrations than the original π₀.₅ model did with 60. This experiment only covered a few official test tasks and cannot yet prove similar gains will generalize across other robots or complex scenarios. But it helped me understand the purpose behind Axis’s data collection: first learn to locate targets, adjust angles, and approach objects in simulation; then use only a small number of real-world interactions to fine-tune contact and grasping on the physical robot. Of course, the trajectories we complete in the Task Hall aren’t directly added to Booster’s training set. They undergo replay, filtering, format conversion, and compatibility checks before being used. Used to focus only on scores and points after completing tasks. Now I pay closer attention to whether my movements were shaky or if my grasping path was clean. After the robot gets a new body, whether we can cut real-world data collection by half really depends on the quality of these trajectories. To try it out, start with Beginner tasks. After passing validation, remember to complete the Base signature in Portfolio: https://t.co/Bo1luv9UkN