No Digg Deeper questions have been answered for this story yet.
SPACE remains robust under dynamic shifts during deployment.
During training, it has only seen the empty box, and we put heavy metal into the box during inference.
SPACE adapts control commands to lift the heavy box, while vanilla policy fails.
It can also execute the policy under different control Hz and controller gains.
📄 Paper: https://arxiv.org/abs/2606.24049 📷 Project page: http://haeone.site/space-website/
Special thanks to awesome collaborators @bkjeon1211 @Suchaeck @jiankimr and adviser @kimin_le2🙏.
P.S. I will be attending ICML 2026 in Seoul, Korea 🇰🇷. Feel free to reach out to me to chat about robot learning! 🧵[n/n]
SPACE improves cross-hardware learning. Even under the same embodiment, different robots exhibit dynamic discrepancies due to hardware wear and tear.
By addressing this, SPACE improves performance in DROID, collected from multiple hardware units of the same Franka robot.
Recent works scale policy learning across different embodiments (Open-X) or different hardware units of the same embodiment (DROID).
🚨However, robot actions are often tied to the dynamics of the data collection robot.
For example, replaying actions from one robot to the other leads to a different trajectory.
🤔 Why? Actions are defined as underlying robot control commands.
Different controller implementations across embodiments result in different actions for the same motion.
Robot dynamics are also affected by hardware wear and tear and manufacturing variability.
❗ This significantly hurts learning & deployment across different robots.
Our framework: SPACE
🔑 SPACE adopts Cartesian state delta as a universal action across different robots.
By predicting robot motion instead of control input as an action, SPACE remains agnostic to data collection dynamics.
🔑 During deployment, Action Adapter continuously converts predicted Cartesian state deltas to robot-specific control commands.
This enables test-time adaptation to different dynamics (eg., object weights, control Hz, and controller gains).
SPACE improves cross-embodiment learning. It improves co-training success rates across Franka Research 3 (FR3) + UR5 robot and FR3 + UMI gripper data, compared to policy predicting control commands. It further enables zero-shot transfer from UR5 to FR3, while the vanilla policy struggles due to the dynamics gap.