Search papers, labs, and topics across Lattice.
This paper introduces StableIDM, a spatio-temporal framework designed to enhance the robustness of Inverse Dynamics Models (IDMs) against manipulator truncation by refining features from visual inputs. StableIDM uses robot-centric masking, Directional Feature Aggregation (DFA) for geometry-aware spatial reasoning, and Temporal Dynamics Refinement (TDR) to smooth action predictions. Experiments on the AgiBot benchmark and real-robot replay demonstrate that StableIDM significantly improves action accuracy, task success, and end-to-end grasp success under severe truncation.
Even with partial observability from manipulator truncation, StableIDM recovers stable action predictions, boosting downstream VLA real-robot success by 17.6%.
Inverse Dynamics Models (IDMs) map visual observations to low-level action commands, serving as central components for data labeling and policy execution in embodied AI. However, their performance degrades severely under manipulator truncation, a common failure mode that makes state recovery ill-posed and leads to unstable control. We present StableIDM, a spatio-temporal framework that refines features from visual inputs to stabilize action predictions under such partial observability. StableIDM integrates three complementary components: (1) auxiliary robot-centric masking to suppress background clutter, (2) Directional Feature Aggregation (DFA) for geometry-aware spatial reasoning, which extracts anisotropic features along directions inferred from the visible arm and (3) Temporal Dynamics Refinement (TDR) to smooth and correct predictions via motion continuity. Extensive evaluations validate our approach: StableIDM improves strict action accuracy by 12.1% under severe truncation on the AgiBot benchmark, and increases average task success by 9.7% in real-robot replay. Moreover, it boosts end-to-end grasp success by 11.5% when decoding video-generated plans, and improves downstream VLA real-robot success by 17.6% when functioning as an automatic annotator. These results demonstrate that StableIDM provides a robust and scalable backbone for both policy execution and data generation in embodied artificial intelligence.
