[10 real demonstrations: 0/20 object contacts. Add 50 more simulation trajectories: 17/20] @KaitoAI X @axisrobotics Axis participation link (referral): https://t.co/hUbWSVsBVR Training robots with new tasks requires significant time and cost to collect demonstration data on actual hardware. You must set up the robot, calibrate the workspace, and manually control it to record dozens or hundreds of trials—and if the object or environment changes, you need entirely new data. Axis and @boosterobotics recently conducted an experiment to see how learning outcomes differ when combining real demonstrations with Axis simulation trajectories using Booster’s dual-arm robot. The tasks used were picking up a pen and picking up an earbud case. Each policy was evaluated separately on whether the robot touched the object, gripped it, and successfully lifted it—all starting from the same Pi0.5 model checkpoint. With Axis Suite, you can briefly film the real workspace and use a pretrained 2D-to-3D model to generate a task-specific digital twin. The generated scene is deployed on Axis Hub, allowing multiple contributors to produce simulation trajectories for the same task—even without physical robots. You can vary object positions, camera viewpoints, lighting, materials, backgrounds, and surrounding objects to perform the same task under multiple conditions. The low-data experiment using earbud cases included 10 real demonstrations. A policy trained on this data failed to touch the target object even once in 20 executions on the actual Booster robot. Keeping the 10 real demonstrations unchanged and adding 50 Axis simulation trajectories resulted in a policy that touched the object 17 out of 20 times. Six of those reached the grasp stage, and two successfully lifted the earbud case. In further evaluations using more data on physical robots, multiple combinations were tested. For the pen task, the real-only policy achieved a lift rate of 35%, while a condition with 67% simulation data achieved 40%. For the earbud case, real-only achieved 30%, and a condition with 50% simulation data achieved 40%. Each condition was executed 20 times on the physical robot, totaling 240 rollouts for the full evaluation. Simulations allow rapid iteration by changing object positions, viewpoints, backgrounds, and other conditions. Real demonstrations capture physical hardware characteristics such as camera properties, actuator response, gripper timing, contact, and execution errors. For both tasks, the highest lift rates were achieved by combining real demonstrations with simulation data tailored to the task. Axis and Booster also developed a model pre-trained specifically for Booster using data collected across multiple tasks. This model learns Booster’s camera setup, two 7-DoF arms, gripper mechanics, action representation, and basic movements before learning new tasks. The original simulation dataset consisted of 44,039 episodes across 95 repositories and 76 language commands. After filtering out static trajectories and recordings where it was unclear which arm was used, 42,046 episodes and 3,249,835 frames remained—equivalent to approximately 180.55 hours at 5Hz. Real demonstrations from Booster with success labels, telemetry, and observations were also included in training. After full-parameter continued pretraining of Pi0.5 on this data, it was fine-tuned for the pen and earbud case tasks. With 30 demonstrations per task, the Booster-specific model successfully executed the sequence of approach → gripper close → lift upward in 14 out of 16 attempts. The original Pi0.5 model achieved only 6/16 under the same 30-demo condition, and improved to 10/16 when demonstrations were increased to 60. Even before task-specific fine-tuning, zero-shot simulation showed reduced average wrist-to-target distance—from 17.27cm to 5.78cm—and reduced unintended arm movement—from 5.04 rad to 0.25 rad. These metrics are results from collaborative experiments by Axis and Booster. No independent replication of these results has been publicly released yet. The low-data evaluation using 10 real demonstrations and 50 simulations was conducted on the physical Booster robot; comparisons of 30 and 60 demonstrations for the Booster-specific model were performed in Isaac Sim. The numbers from these two evaluations cannot be combined into a single metric for physical robot performance. Simulation trajectories generated on Axis Hub were used to vary object placement and visual conditions; real Booster demonstrations contributed physical hardware characteristics such as contact and execution behavior to training. Data aggregated across multiple tasks also contributed to building an initial model familiar with Booster’s body and control methodology.Ang Trajectory na nakalap sa browser ay ginamit talaga sa pag-e-evaluate ng pisikal na robot ng Booster at sa pagkatuto ng unang modelo. #PhysicalAI #RobotLearning