Pinned Tweet
Axis is officially LIVE on @base. 🔵
Axis is scaling Physical AI for the real world, contributed by everyone.
You can control robots in a virtual world, generate training data at scale, and help build the brain behind tomorrow's robots. All from browser. No hardware needed.
Start building robotics intelligence today: hub.axisrobotics.ai
Robotic skills, distilled from 11,199 real manipulation tasks and 14.6 TB of open data.
Introducing the AXIS Skills Atlas: every skill mapped across embodiments, each number opening the real tasks behind it.
skills.axisrobotics.ai/
6/7 Task explorer
Search by task name, or filter by:
• Skill
• Robot
• Scene: kitchen, workshop, office, home and more
• Difficulty: 1-5 stars
Watch demos, build a shortlist and export it as CSV.
7/7 Built on open data
Our open Franka task datasets are in LeRobot v3.0 format: Parquet + MP4 at 15 fps, third-person and wrist cameras, joint states, actions and end-effector pose per frame.
14.6 TB of data is open on Hugging Face, with a sample task ready to download.
Need a skill, custom task data or a commercial licence? Tell us your robot, tasks and eval goal: [email protected]
Every trajectory now earns $AXIS.
Points are done. From Saturday 14 November, every verified, signed trajectory earns $AXIS, and you can claim it every week.
Explore now: hub.axisrobotics.ai/
How it works ⬇️
5/6 Claim weekly
The amount shown the moment you sign is the amount you get.
Rewards settle every Saturday at 10:00 UTC. Open Portfolio-Rewards, and claim straight to your wallet on Base.
Miss a week and nothing is lost. Unclaimed $AXIS never expires.
Live: Sat 14 Nov, 10:00 UTC.
First claim: Sat 21 Nov, 10:00 UTC.
We’ve been using ODE Policy for visual policy distillation, with strong early results. Learning how actions evolve over time gives the policy a natural way to capture the continuity of robot motion.
In the video, our Franka robot learns to thread a nut onto a bolt in simulation, reproducing the task from SGS.
Next, we plan to use ODE Policy to distill a wider range of assembly skills into policies that act from camera images, from fitting gears and inserting pegs to joining connectors.
🚀 Nature does not move in discrete jumps. Why should robot policies?
Super excited to release ODE Policy — perhaps our most exciting breakthrough in my lab this year! 🤖🔥
Think about how humans reach for a cup. Our movements evolve continuously, governed by underlying dynamics. Yet, most robot policies — from Diffusion and Flow Matching to VLAs — learn to predict explicit actions.
What if robots learned how actions evolve, instead of predicting actions?
💡 That's ODE Policy for Embodied AI.
We use a neural Ordinary Differential Equation (ODE) to learn continuous action dynamics, rather than directly generating actions.
One simple idea. Five types of robots. Multiple learning paradigms.
✅ Single-arm manipulation
✅ Dual-arm manipulation
✅ Dexterous hands
✅ Humanoid whole-body control
✅ Aerial robots
And it goes far beyond imitation learning:
⚡ VLA (π0.5): 96.0% vs. 75.5% success after one epoch.
🌍 4.5B World Action Model: 82.6% vs. 24.6% success, with ~4× faster action generation.
🏃 RL: More efficient humanoid whole-body motion learning.
For years, we've focused on scaling robot data, models, and compute.
But what if the next fundamental breakthrough comes from rethinking WHAT robots learn, rather than just HOW MUCH they learn?
Huge congratulations to Jindou Jia and our amazing team! 👏
📄 Paper: arxiv.org/abs/2611.01729
🌐 Project: ntumars.github.io/ODE/
#ODEPolicy #EmbodiedAI #PhysicalAI #RobotLearning #WorldModels #VLA #MARSLab
How should you choose simulation demos for co-training when real robot data is scarce?
Introducing TUCO: Curating Simulation Demonstrations for Sim-to-Real Robot Policy Co-Training, the first systematic study of data curation for sim-to-real robot policy co-training.
With just 10 real demos and 600 curated sim demos per task, TUCO hits 85.8% real-robot success, versus 25.8% for the full 1,200 pool or a random 600, and sets SOTA across all three transfer settings.
A collaboration with researchers from @Stanford @NTUsg @CarnegieMellon @FudanUniversity @UCBerkeley @sjtu1896.
Project page: tuco-curation.github.io/
Details below ⬇️
5/6 Simulation Experiments
We validate the criterion in two simulated settings with the same protocol, averaging over three seeds.
Single-simulator (RoboMimic). Source and target share one simulator, which isolates the curation criterion from any domain gap. On Lift, Square, and Transport, TUCO is best or tied-best in 18 of 25 task-budget settings, and matches all-data training on Lift using only 25% of the data.
Sim-to-sim (IsaacSim to MuJoCo). Demos are generated in one simulator and evaluated in another, a controlled stand-in for the sim-to-real gap. TUCO is best in 15 of 18 settings and above random selection in all 18. The full IsaacSim pool even hurts CupCake, while a curated subset gives the best result on every task.
1
1
811
6/6
When real data is scarce, careful selection drives the value of sim data more than raw volume. Influence-based utility and coverage offer a closed-loop way to pick demos that genuinely transfer to real robots.
Project page: tuco-curation.github.io/
Paper: arxiv.org/abs/2610.05407
Code: github.com/Z-yiwei/tuco
1
712






