2026

Plan-Conditioned Imitation for Robust Object Retrieval
under Self-Occlusion in Dense Clutter

Kowndinya Boyalakuntla1Ajinkya Pawar2Abdeslam Boularias1Jingjin Yu1

1 Rutgers University2 Indian Institute of Technology Bombay.

Teacher Rollouts for Adaptive Closed-loop Execution

A robot’s arm blocks its own view. TRACE uses one initial plan, memory,
and partial feedback to retrieve a target object from dense clutter.

Watch the film 2:59 Read the paper Code Data

Overview

Supplementary video · 2:59
TRACE IN ACTIONUR5e · Real robot experiments

Target retrieval in dense clutter under self-occlusion.

TranscriptDownload video

INSIDE TRACE

Training & inference

A fixed teacher rollout provides local plan context. Partial object observations and robot state update a recurrent student, which selects pushing actions before a final graspability check. View full resolution
One fixed teacher rollout guides a recurrent student using partial observations and memory.

On the real robot, we estimate object poses from instance masks produced by Mask R-CNN (He et al., ICCV 2017).

ON THE REAL ROBOT

Demonstrations

Choose a method and scene.

Outcome: Success

Scene 01

90.0% success67.3 s total2.9× faster than Online Teacher, with no sensing retractions during pushing.

Selected videos; statistics cover the full hardware benchmark: 20 scenes × 2 trials per method. Total time includes initialization/planning, execution, and grasp/lift.

Citation

@misc{boyalakuntla2026trace,
  title = {Plan-Conditioned Imitation for Robust Object Retrieval under Self-Occlusion in Dense Clutter},
  author = {Boyalakuntla, Kowndinya and Pawar, Ajinkya and Boularias, Abdeslam and Yu, Jingjin},
  year = {2026},
  url = {https://trace-retrieval.github.io/}
}