Back to research
Research Paper
Spatial Heatmap Representations for Few-Shot Manipulation
A technical validation record for preserving spatial structure across vision, language and action in FAM.
Published March 5, 2026Updated June 15, 2026
FAMSpatial RepresentationFew-Shot Learning
Research question
Can spatial structure be preserved more directly than a compressed one-dimensional representation while retaining an end-to-end manipulation policy?
Approach
FAM aligns visual, language and action information through two-dimensional and three-dimensional heatmap representations. The representation exposes task-relevant keypoints, object relationships and motion cues without spatial information loss.
Evaluation
The evaluation framework includes RLBench, peg insertion, drawer opening, lighting variation, background interference and obstacle interference.