[10 real demonstrations: 0/20 object contacts. Add 50 simulation trajectories → 17/20] @KaitoAI X @axisrobotics Axis participation link (referral): https://t.co/hUbWSVsBVR Collecting demonstration data on actual hardware to teach robots new tasks requires significant time and cost. You must set up the robot, calibrate the workspace, and manually teleoperate it to record dozens or even hundreds of trials—and if the object or environment changes, you need entirely new data. Axis and @boosterobotics recently conducted an experiment to evaluate how learning outcomes change 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. Performance was evaluated separately on whether the robot touched the object, successfully gripped it, and actually lifted it. All policies started 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. This scene is deployed on Axis Hub, enabling multiple contributors to generate 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 diverse conditions. In the low-data experiment using earbud cases, 10 real demonstrations were used. The policy trained on this data failed to make contact with the target object in any of the 20 trials on the actual Booster robot. Keeping the 10 real demonstrations unchanged and adding 50 Axis simulation trajectories resulted in a policy that made contact with the object in 17 out of 20 trials. Six of those reached the grasp phase, and two successfully lifted the earbud case. In further evaluations using more data on the physical robot, 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 50% simulation condition achieved 40%. Each condition was evaluated with 20 trials on the physical robot, totaling 240 rollouts. Simulations allow rapid iteration by changing object positions, viewpoints, backgrounds, and other conditions. Real demonstrations capture physical nuances such as camera characteristics, actuator response, gripper timing, contact dynamics, 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 trained a model initialized specifically for Booster using data collected across multiple tasks—a model pre-trained on Booster’s camera setup, two 7-DoF arms, grippers, action representations, and basic movements before learning new tasks. The original simulation dataset comprised 44,039 episodes across 95 repositories and 76 language commands. After filtering out static trajectories and records where it was unclear which arm was used, 42,046 episodes and 3,249,835 frames remained—equivalent to approximately 180.55 hours at 5 Hz. Real Booster demonstrations with success labels, telemetry, and observations were also included in training. After full-parameter continued pretraining of Pi0.5 on this dataset, the model 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 in 14 out of 16 trials. The original Pi0.5 model achieved only 6/16 under the same conditions and improved to 10/16 when demonstrations were increased to 60. Even before fine-tuning, zero-shot simulation performance improved: average wrist-to-target distance decreased from 17.27 cm to 5.78 cm, and unintended arm movement dropped from 5.04 rad to 0.25 rad. These metrics reflect results from Axis and Booster’s collaborative experiments. No independent replication of these exact results has been publicly reported yet. The low-data evaluation using 10 real demonstrations and 50 simulation trajectories was conducted on the physical Booster robot; comparisons of 30 and 60 demonstrations for the Booster-specific model were performed in Isaac Sim. These two sets of numbers cannot be directly combined into a single physical robot performance metric. Simulation trajectories generated via Axis Hub were used to diversify object placements and visual conditions; real Booster demonstrations contributed physical properties such as contact dynamics and execution behavior to training. The curated dataset from multiple tasks also contributed to building an initial model familiar with Booster’s body and control methodology.Trajectories collected in the browser were actually used for physical robot evaluation and initial model training of the Booster. #PhysicalAI #RobotLearning