Axis Robotics Releases AXIS, a Browser-Based Data Engine With 207 Robot-Manipulation Tasks and 50,129 Trajectories

AXIS separates demonstration collection from backend processing. Contributors teleoperate a Franka Research 3 with a parallel-jaw gripper in a MuJoCo WebAssembly front end, using keyboard, mouse, virtual joystick or gamepad; no local GPU or robot is required. Physics stepping and Three.js rendering run outside the React UI thread, and the backend handles the heavy workload with 8x RTX 4090 GPUs for rendering and 8x A100 GPUs for training and evaluation.

Tasks are generated rather than hand-authored. TaskGen turns a language instruction into task, scene and object configurations, retrieves or generates meshes through an image-to-3D pipeline, rescales them, and proposes a 2.5D layout. A layout supervisor validates the scene and can relocate, reorient or regenerate objects if constraints fail. Each task includes a structured success checker that the backend reruns rather than trusting the front-end success flag.

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, and third-view plus wrist RGB-D observations. The paper underlying the release credits more than 70,000 community members with contributions.

Cleaning is treated as a production stage. Samples with joint variation below 5e-3 are dropped as static, a Savitzky-Golay filter smooths continuous motion, and cubic splines resample the 6 Hz to 8 Hz output of the web interface to a 20 Hz target. The reported tradeoff is that mean acceleration falls from 1.3539 to 0.4885, mean jerk falls from 11.5899 to 2.2243, and replay success falls from 100% to 86.2%. Cleaned episodes are replayed in IsaacSim with physics stepping disabled, so the verified trajectory remains authoritative while scenes, cameras, materials and lights are randomized. Output is 256x256 ray-traced RGB from a fixed third-view camera and a wrist camera, with depth off by default.

To evaluate AXIS, the researchers started every condition from the released π0.5 checkpoint, a PaliGemma Gemma-2B backbone with a Gemma-300M action expert, and followed a common protocol that could optionally continue pretraining on a simulation corpus for 100,000 steps before 30,000 steps of LIBERO post-training. Pretraining used full-model flow-matching loss over 10-step action chunks and no LoRA.

On LIBERO-Plus, π0.5 with AXIS-100% scored 88.8 overall, compared with 83.9 for vanilla π0.5 and 57.5 for a RoboCasa365 control matched on trajectory count. The paper reports improvements of 5.8% and 37.3%; both are relative figures normalized by the 83.9 baseline, so the point gaps of 4.9 and 31.3 are the cleaner read. Aggregate results rose from 84.7 to 85.7 to 88.8 across the 25%, 50% and 100% snapshots of AXIS data.

The gains varied by perturbation axis. The largest improvements were 13.7 points for sensor noise and 11.3 for camera perturbations, with smaller gains for background, robot pose and layout; light and language perturbations regressed by 1.7 and 1.3. The report also notes that camera performance dipped to 68.8 with the AXIS-50% snapshot, below the vanilla baseline of 72.5, before recovering.

According to the report, the AXIS training code is public as a patch layer over OpenPI, and the teleoperation platform is live in any browser. The Hugging Face dataset is gated at 2.36 TB and restricted to non-commercial academic use; no policy checkpoints have been released.