Our #CoRL2026 paper "Hardware-as-Regularization" now has a project page!
A ~$100 3D-printed spring joint physically regularizes leader-arm posture for more consistent IL demos.
by SB Intuitions Robotics Team
https://t.co/HUFZtNa5Jg
I was doing this demo. In the beginning, I just prompted to connect to the robot and it figured out how to connect through ssh, and when I asked to move it's neck, it just went through the codebase and understood how to command each joint.
Our paper at #IROS2026 (1/5)
"Stereo Multistage Spatial Attention for Real-Time Mobile Manipulation Under Visual Scale Variation and Disturbances" (Xianbo Cai et al.)
4 tasks incl. base motion from only 54 demos/task — 85% avg success, ~30 Hz inference on onboard CPU only.
(3/5)
Method: we propose MSARNN — stereo multistage spatial attention (MSA), which extracts task-relevant spatial attention points from stereo images, combined with a hierarchical-LSTM-based deep predictive learning module for closed-loop motion generation.
(2/5)
Problem: in unstructured homes there are no external cameras — the robot must move while working. Its onboard camera viewpoint shifts as it moves, drastically changing the apparent scale of target objects, which destabilizes vision-based motion generation.
(4/4)
Results:
10 Hz inference on Jetson Orin NX; ACT with 10-step history: 0.362 s (~7x slower)
+15-25% success vs Diffusion Policy; 45% vs ~20% (DP) at 2x unseen speeds
generalizes to untaught random motion patterns & speeds
Our paper at #IROS2026 (1/4)
"Energy-Score Stabilized Attentive Mixture-Density Policy: A Lightweight Approach to Multimodal Robot Motion" (Sachiya Fujita et al. )
Left or right? When to grasp? Multimodality is hardest with little data. Attention + GMM policy via Energy Score.
(3/4)
Method: spatial attention + Gaussian mixture outputs trained with the Energy Score (sum of 1D CRPS) — no NLL singularities, stable learning. LSTM-based long-term context reads the moving object's phase & velocity from the past few seconds.
(2/4)
Problem: with several correct actions per situation (obstacle avoidance side, grasp timing), deterministic regression averages the modes, and NLL-trained MDNs suffer variance-collapse singularities. Small-data regimes make this especially severe.