Get Started
Home
Topics
Search
Library
Research questionHow can object-centric world models predict manipulation-induced changes without accumulating errors that derail planning?In tabletop manipulation, most objects may remain stationary while contacts change only a few poses. Predicting the full scene at every step can waste capacity and allow small errors to accumulate, making model-based planning unreliable.
AI
Machine Learning
Research Paper
Robotics
Latest papersRecent research connected to this question, newest first.Modeling What Changes: Sparse, Residual World Models for Object-Centric ManipulationEvidence comes from MuJoCo tabletop pushing scenes containing 3–8 objects. The tested sparse/residual model uses per-object change detection and predicts pose deltas only for flagged objects; it improved next-state pose accuracy, change detection, parameter and data efficiency, object-count transfer, and rollout stability relative to a dense multilayer perceptron. In sampling-based planning, the sparse model trained on planner-visited states achieved 0.23 ± 0.06 success across three seeds, while the dense model remained at zero; a true-simulator oracle solved the task. The evidence does not establish performance beyond this simulator and task setting.research paper · Sep 2, 2026
Related questions
Which architectural, training, and planning choices make joint-embedding predictive world models generalize to new physical tasks?How can robots safely learn dynamic manipulation skills online despite sim-to-real mismatch?How can long-horizon indoor mobile manipulators track embodiment state and repair plans without dense scene maps?How can pretrained vision-language-action models reliably perform contact-rich manipulation when goals, scenes, and contacts change?