AXIS Lifts π0.5 to 88.8 on LIBERO-Plus With 50,129 Browser-Collected Trajectories

Axis Robotics, in collaboration with researchers from UC Berkeley, Georgia Tech, and NTU, has released AXIS, a robotics demonstration platform that relocates teleoperation into the browser while delegating compute-heavy workloads to backend GPU clusters. The dataset snapshot on Hugging Face contains 207 manipulation tasks, 50,129 episodes, and more than 60,000 task and scene variants across seven scene categories, totalling 2.36 TB and gated to non-commercial academic use. Training code ships as a patch layer over OpenPI and is publicly available; policy checkpoints are not released.

The bottleneck AXIS targets is structural: conventional manipulation datasets are collected on lab hardware by expert operators, processed offline, and published as fixed benchmarks. AXIS treats the dataset as a continuously expanding artifact, with more than 70,000 community members credited as contributors in the paper.

Browser-GPU Architecture and the TaskGen Pipeline

The system's core design choice is an explicit asymmetry between what runs in the browser and what runs on backend hardware. Contributors teleoperate a Franka Research 3 with a parallel-jaw gripper inside a MuJoCo WebAssembly frontend, using keyboard, mouse, virtual joystick, or gamepad — no local GPU, no physical robot. Physics stepping and Three.js rendering execute off the React UI thread, keeping logged state-action samples synchronized with simulator state rather than with interface latency. Rendering runs on 8× RTX 4090 GPUs; training and evaluation run on 8× A100 GPUs.

Task generation is automated rather than hand-authored. The TaskGen module decomposes a natural-language instruction into task, scene, and object configurations, retrieves or synthesizes meshes through an image-to-3D pipeline, rescales them to physically plausible dimensions, and proposes a 2.5D layout. A layout supervisor validates each instantiated scene and relocates, reorients, or regenerates objects when geometric or physical constraints fail. Every task ships with a structured success checker that the backend re-executes independently, rather than trusting the success flag produced by the frontend.

This separation of validation authority from the contributor-facing interface echoes the infrastructure governance challenges that complicate agentic deployment at scale.

Trajectory Refinement and Augmentation

Web teleoperation logs at 6–8 Hz and carries jitter and idle pauses. AXIS runs a three-stage cleaning pass: episodes with joint variation below 5e-3 are discarded as static; a Savitzky-Golay filter with window 15 and polynomial order 3 smooths continuous motion (discrete gripper transitions can bypass this pass); cubic splines resample trajectories to a 20 Hz target.

From Table 1 of the paper: 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%. Cleaned episodes are then replayed in IsaacSim from packed simulator state with physics stepping disabled, preserving the verified trajectory as authoritative while randomizing scene, camera pose, materials, and lighting. Output is 256×256 ray-traced RGB from a fixed third-view camera and a wrist camera; depth is disabled by default.

LIBERO-Plus Benchmark Results

Every evaluated condition initializes from the released π0.5 checkpoint — a PaliGemma Gemma-2B backbone paired with a Gemma-300M action expert — optionally runs continual pretraining on a simulation corpus, then fine-tunes on LIBERO-Plus under identical hyperparameters. Pretraining is full-model with no LoRA, using a flow-matching loss over 10-step action chunks for 100,000 steps, followed by 30,000 steps of LIBERO post-training.

Aggregate scaling is monotonic across the 25%, 50%, and 100% snapshots, but per-axis results are uneven. Sensor Noise gains 13.7 points and Camera gains 11.3 points against the vanilla baseline — both axes where the augmentation pipeline actively randomizes. Background gains 3.7 points, Robot Pose 3.8, and Layout 2.6. Light regresses 1.7 points and Language regresses 1.3 points. Camera additionally dips to 68.8 at AXIS-50%, below the 72.5 vanilla baseline, before recovering at full data volume. Real-robot evaluation in the paper is qualitative only.

The RoboCasa365 control result — 57.5 on the same volume-matched trajectory count — isolates the data quality argument from the data volume argument: the AXIS pipeline's domain-randomized augmentation accounts for the gap, not simulation scale alone. This aligns with the broader pattern that architectural specificity outperforms raw GPU scaling when the inductive structure matches the target task distribution.

The 31.3-point gap between AXIS-100% and the RoboCasa365 control gives the crowd-sourced browser teleoperation approach its sharpest empirical support so far. The Language and Light regressions signal that browser-collected data still carries domain gaps the current augmentation pipeline does not fully close.