Axis Robotics Releases AXIS: A Browser-Based Data Engine With 207 Robot Manipulation Tasks and 50,129 Trajectories
207 tasks, 50K trajectories: how Axis moved robot data collection from labs to browsers—and closed a decade-old bottleneck.
Why it matters
Axis decouples robot demonstration collection from hardware by moving it to the browser with backend GPU processing. This unlocks large-scale trajectory datasets—a historically scarce input for embodied AI training—and shows measurable gains: π0.5 jumps to 88.8 on LIBERO-Plus versus 57.5 for volume-matched baselines. A direct attack on the data scarcity that has held back robot learning.
The key facts
6 to know207 robot manipulation tasks in AXIS dataset
50,129 verified Franka trajectories collected
π0.5 performance: 83.9 → 88.8 on LIBERO-Plus after continual pretraining
RoboCasa365 control baseline: 57.5 (volume-matched)
Data collection infrastructure: browser frontend, backend GPU processing
Removes hardware lock-in from trajectory collection
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Robot datasets have grown far slower than the models trained on them, mostly because collection stays locked to lab hardware. AXIS moves demonstration collection into a web browser and pushes everything expensive to backend GPUs. The result is 207 tasks and 50,129 verified Franka trajectories, and…