AXIS: A Browser-Based Robot Data Engine With 50,129 Trajectories

Axis Robotics, together with researchers from UC Berkeley, Georgia Tech and NTU, released AXIS, a system that moves robot manipulation data collection into the browser. Contributors teleoperate a Franka Research 3 with a parallel-jaw gripper inside a MuJoCo WebAssembly frontend using keyboard, mouse, virtual joystick or gamepad. Physics stepping and Three.js rendering run off the React UI thread, keeping logged state-action samples aligned with the simulator. The heavy work sits on servers: rendering on 8x RTX 4090 GPUs, training and evaluation on 8x A100 GPUs.

The released snapshot contains 207 tasks, 50,129 episodes and more than 60,000 task or scene variants across seven scene categories. Each trajectory carries task metadata, embodiment, simulator version, robot and object states, actions, success labels, plus third-view and wrist RGB-D observations. The paper credits more than 70,000 community members.

Generated tasks, verified replays

A component called TaskGen decomposes a language instruction into task, scene and object configs, retrieves or generates meshes through an image-to-3D pipeline, rescales them to plausible physical size, then proposes a 2.5D layout. A layout supervisor relocates, reorients or regenerates objects when constraints fail. Every task ships with a structured success checker that the backend re-runs instead of trusting the frontend flag.

Cleaning drops samples with joint variation below 5e-3 as static, applies a Savitzky-Golay filter with window 15 and polynomial order 3, and resamples the interface's 6 Hz to 8 Hz output up to 20 Hz with cubic splines. Mean acceleration falls from 1.3539 to 0.4885 and mean jerk from 11.5899 to 2.2243, while replay success drops from 100% to 86.2%.

Training starts from the released π0.5 checkpoint. On LIBERO-Plus, π0.5 plus AXIS-100% reaches 88.8 overall against 83.9 for vanilla π0.5 and 57.5 for a RoboCasa365 control matched on trajectory count. Aggregate scores move 84.7, 85.7, 88.8 across the 25%, 50% and 100% snapshots. The largest per-axis gains are Sensor Noise +13.7 and Camera +11.3, while Light and Language regress by 1.7 and 1.3. The 2.36 TB Hugging Face dataset is gated and restricted to non-commercial academic use, and no policy checkpoints were released.

Full details are in the MarkTechPost report.