Abstract
Visual inverse dynamics models decode low-level robot actions from camera observations and are increasingly used for robot replay and visual-plan execution. However, in single-view manipulation, the robot arm often moves partly outside the camera field of view, making the inverse problem ill-posed and causing unstable action predictions. We present StableIDM, a causal spatio-temporal inverse dynamics model designed for manipulator truncation. StableIDM combines robotcentric masking to suppress non-robot shortcuts, Directional Feature Aggregation to preserve anisotropic evidence across fixed image-space directions, and Temporal Dynamics Refinement to aggregate evidence available in a short causal history and smooth the decoded action. Within the evaluated fixed-camera, sameembodiment setting, a single-seed matched comparison on the AgiBot truncation benchmark shows that StableIDM improves low-occupancy strict accuracy by 5.00 percentage points over a masked causal temporal baseline. In modest-scale physical studies, the same decoder achieves numerically higher openloop replay and visual-plan grasp success. Additional analyses characterize mask quality and deployment cost.
| Original language | English |
|---|---|
| Journal | IEEE Robotics and Automation Letters |
| DOIs | |
| Publication status | Accepted/In press - 2026 |
| Externally published | Yes |
Keywords
- Deep Learning in Grasping and Manipulation
- Machine Learning for Robot Control
- Perception for Grasping and Manipulation
Fingerprint
Dive into the research topics of 'StableIDM: Robust Inverse Dynamics for Robot Manipulation Under Manipulator Truncation'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver