So AXIS runs every trajectory through a 4-stage cleaning pipeline: quality filter, hesitation removal, Savitzky Golay smoothing, cubic spline resampling.
Result on 16 LIBERO benchmark tasks: motion jerk down 47%, acceleration spikes down 54%, while keeping over 95% of the data.
When you teleop a robot in your browser, your hand shakes, hesitates and overshoots. Feed that raw jitter to a real robot and it amplifies into unsafe, jerky motion.
@axisrobotics@AxisRoboticsID Every turn compounds: more data, smarter robots, bigger network.
200K+ trajectories and 8,900+ nodes isn't the ceiling. It's momentum.
AXIS isn't a tool. It's a flywheel.
Demo in a browser → trajectories cleaned → 1 demo becomes thousands in IsaacSim → VLA models train → real robots deploy → TaskGen spawns new tasks → contributors come back for more.
@axisrobotics Leveling up 🔥
With the introduction of Long-Horizon (LH) and Multi-Embodiment tasks, Axis is moving beyond simple robotic actions toward more complex, real-world scenarios.
This means:
✅ richer training data
✅ more diverse robot interactions
✅ better adaptability across different robot forms
The goal isn't just collecting more data, it's building smarter and more valuable data for the future of Physical AI. 🚀