13 papers this week in Robotics & embodied AI12 active threadsbusiest: Alignment Training Dynamics

Robotics & embodied AI research

Robotics & embodied AI

Robots can regain lost success after camera shifts without retraining

Lyu and colleagues show that transforming a moved camera’s images back toward the training viewpoint can substantially recover robot task performance without retraining the action model. Across roughly 300,000 simulated episodes, their adapter improved four robot policies, with a crucial distinction: it reconstructed the surroundings visually but rendered the robot arm from its known geometry and joint positions.

3 days ago
Robotics & embodied AI

Robots can improve actions by changing randomness used for video prediction

Liu et al. show that pretrained robot policies can perform tasks more reliably by changing the random inputs from which they generate predictions, without updating the policies themselves. Across ten model-task pairs, their video-noise steering method raised average success by 26.2 percentage points over the base policies, compared with 7.1 points for an existing action-noise steering method.

3 days ago
Robotics & embodied AI

Tactile sensors trained on wrench data work across different hardware

The authors collected over 100 million paired tactile observations and wrench measurements from 24 sensors across four sensing families. They trained a unified encoder to reconstruct wrenches, producing a shared feature space that generalizes to unseen sensors. The approach reduced force estimation error by 17.1% and improved manipulation success rates by up to 21.7 percentage points over baselines.

4 days ago