PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Kowndinya Boyalakuntla, Ajinkya Pawar, Abdeslam Boularias, Jingjin Yu

arXiv:2609.38857v1cs.RO

TL;DR

Retrieval in dense clutter requires rearranging objects even as the manipulator can occlude them, making partial-observation feedback important. TRACE conditions a recurrent student on a fixed teacher rollout and refines it with teacher labels on student-visited states. It reports improved retrieval outcomes in simulation and on hardware, alongside a benefit from student-state supervision at matched label budgets.

  • Problem

    During clutter rearrangement, the manipulator can occlude objects while contact continues to move them, and arm withdrawals to restore visibility interrupt manipulation.

    Source: §I · p. 1

  • Method

    TRACE uses a fixed teacher-generated rollout as predictive context for a recurrent student acting from partial observations, with DAgger supervision on student-visited states.

    Source: §I · p. 1, D. Learning from Privileged Supervision · p. 5

  • Results

    At a matched 26,373-label budget, success is 87.8% for student-state supervision versus 66.7% for expert-only cloning; on hardware, TRACE reaches 90.0% success versus 47.5% for TEACHER REPLAY.

    Source: C. Effect of Student-State Supervision · p. 6, E. Real-Robot Evaluation · p. 7

  • Takeaways & Limitations

    The nominal rollout supports feedback under partial observability, while matched-label comparisons show student-visited-state supervision improves retrieval reliability beyond adding teacher demonstrations alone.

    Source: §VI · p. 8

    The current evidence is limited to planar clutter with known object footprints and an initially visible scene.

    Source: §VI · p. 8

Abstract

from arXiv · show

Retrieving objects from dense clutter requires rearrangement during which the manipulator can occlude objects while moving them. Repeated arm withdrawals to restore visibility interrupt execution. We introduce TRACE, a plan-conditioned imitation framework for retrieval under self-occlusion. A single unoccluded observation initializes a digital twin, where a privileged teacher generates a fixed nominal rollout. A recurrent student combines local rollout context, partial object observations, and proprioception to select actions that can correct deviations from the prediction. Behavior cloning initializes the student; DAgger refines it with teacher labels on student-visited states. The rollout remains fixed throughout execution, so the deployed student needs neither online teacher queries nor additional simulator rollouts during pushing. On 511 simulation test scenes, TRACE achieves 90.7% success versus 43.4% for nominal replay and 96.7% for the privileged closed-loop teacher. At a matched 26,373-label budget, student-state supervision achieves 87.8% versus 66.7% for expert-only cloning, demonstrating gains beyond additional labels. On a UR5e, TRACE achieves 90.0% success versus 95.0% for the closed-loop teacher, while reducing total execution time from 192.7 s to 67.3 s. It avoids the teacher's 16.8 sensing-related arm retractions per trial during pushing, retaining a final withdrawal for graspability evaluation. Code and data will be released at: https://trace-retrieval.github.io.

I. INTRODUCTION

TRACE addresses retrieval under manipulator-induced self-occlusion by combining a fixed teacher-generated plan with recurrent control from partial visual feedback. It also uses teacher labels on student-visited states to refine behavior cloning.

I. INTRODUCTION

  • Self-occlusion can hide objects while contact continues to change their poses, whereas arm retractions restore visibility at the cost of interrupting manipulation.Replaying a predicted sequence cannot correct for contact uncertainty or execution error.Source: §I · p. 1
  • TRACE uses a fixed, scene-specific teacher rollout as predictive context while a recurrent student acts from partial observations and can correct deviations.The rollout is generated from one unoccluded observation in a digital twin; it is not replayed as a sequence of required actions.Source: §I · pp. 1–2, Figure 2 · p. 2
  • DAgger refines behavior cloning with teacher labels on student-visited states, including configurations absent from teacher-controlled demonstrations.Training includes simulated self-occlusion, detection dropout, multi-step observation blackouts, and planning–execution mismatch.Source: §I · p. 2
  • The paper evaluates TRACE in simulation and on a real robot, reporting improved retrieval success over nominal-plan replay and reduced execution time relative to a complete-state teacher.The reported time reduction is attributed to avoiding repeated arm withdrawals for sensing during pushing.Source: §I · p. 2

II. RELATED WORK

Prior retrieval work includes predictive planning and reactive policies, while TRACE is distinguished by rollout-conditioned recurrent feedback under partial observation. The problem formulation specifies planar clutter retrieval with known object footprints, an initially complete scene observation, and partial observations during manipulation.

A. Object Retrieval through Non-Prehensile Rearrangement

  • TRACE uses one pre-execution rollout as predictive context, unlike prior predictive methods that search over outcomes or accelerate long-horizon planning through simulation.It also differs from sequential pre-grasp pushing without retraction by using rollout-conditioned recurrent feedback and explicit missing-object observations.Source: A. Object Retrieval through… · p. 2

B. Privileged Learning under Partial Observations

  • TRACE combines recurrent observation history with a nominal rollout from the initial scene; privileged state supplies training labels and pre-execution context, while deployment uses partial observations and the fixed rollout.Related work includes privileged teacher distillation and memory-augmented student-teacher learning under imperfect detections.Source: B. Privileged Learning under Partial… · p. 2

C. Imitation on Learner States and Nominal Guidance

  • TRACE matches label budgets to isolate learner-state coverage and predicts the complete next motion primitive rather than adding a learned residual to a conventional controller.DAgger labels learner-induced states; in rearrangement, a different push changes subsequent contacts and observations.Source: C. Imitation on Learner States and… · p. 2

III. PROBLEM FORMULATION

  • The task is target retrieval from dense planar clutter through non-prehensile pushes, with known object footprints and a complete initial scene observation.The formulation models execution as a finite-horizon discounted POMDP, with 16 discrete planar motion primitives and observations that may omit objects under manipulator self-occlusion.Source: §III · pp. 2–3

IV. METHOD

TRACE generates a fixed nominal rollout from an initial scene estimate, then uses it as local context for a recurrent student acting from partial observations. The rollout serves as predictive context rather than commands to replay.

IV. METHOD

  • A frozen privileged teacher is rolled out once in a digital twin initialized from the estimated initial scene.This produces a scene-specific nominal trajectory for the student’s execution.Source: §IV · p. 4
  • A GRU-based student executes from partial observations while conditioning on local context extracted from the fixed rollout.The student can depart from the predicted execution as the physical scene evolves.Source: §IV · p. 4
  • The nominal trajectory provides predictive context rather than commands that the student must replay.The rollout is fixed while the student responds to the evolving physical scene.Source: §IV · p. 4

A. Learning a Privileged Retrieval Teacher

A complete-state teacher uses a permutation-invariant scene representation to choose among 16 motion primitives, trained with a graspability-based reward.

A. Learning a Privileged Retrieval Teacher

  • The privileged teacher observes complete task geometry, all object poses, and the end-effector state, then selects among the same 16 motion primitives used at deployment.Each object is represented by end-effector-relative planar geometry and a target indicator.Source: A. Learning a Privileged Retrieval… · p. 4
  • The scene encoder pools shared object-MLP features by mean and componentwise maximum, ensuring permutation invariance without a universality claim.The teacher scene representation is formed from privileged object tokens.Source: A. Learning a Privileged Retrieval… · p. 4
  • The teacher is optimized with a graspability-based reward that includes a step cost and workspace-violation penalty, then frozen for discrete action selection.The reward uses graspability confidence and potential differences; policy invariance is not assumed.Source: A. Learning a Privileged Retrieval… · p. 4

B. Nominal Rollout and Local Plan Context

A one-time rollout of the frozen teacher provides a fixed, scene-specific reference, while local geometric context is updated relative to the executing end effector.

B. Nominal Rollout and Local Plan Context

  • Before execution, the initial digital-twin state is used to roll out the frozen teacher once, retaining a nominal trajectory of end-effector and predicted object positions.The trajectory remains fixed throughout the episode.Source: B. Nominal Rollout and Local Plan… · p. 4
  • At each decision, local plan context samples four rollout positions, clamping requested indices beyond the stored trajectory to its final sample.The context also includes rollout progress and a nominal-rollout-exhaustion indicator.Source: B. Nominal Rollout and Local Plan… · p. 4
  • Nominal positions are expressed relative to the current measured end-effector position, so the rollout stays fixed while its geometric context changes with physical execution.The student receives geometric plan context rather than nominal primitive identities.Source: B. Nominal Rollout and Local Plan… · p. 4

C. Plan-Conditioned Recurrent Student

TRACE's recurrent student combines partial object observations, proprioception, and local rollout context; behavior cloning initializes it, and DAgger adds teacher labels on student-visited states.

C. Plan-Conditioned Recurrent Student

  • Visible objects contribute end-effector-relative geometry, while unavailable geometry is zeroed and visibility, observation age, and target identity remain encoded.The structured observation provides one token for each tracked object.Source: C. Plan-Conditioned Recurrent Student · p. 4
  • The recurrent student fuses scene features, end-effector features, local plan context, and the previous action, using a GRU to carry temporal context into action selection.The recurrent state is not trained to explicitly reconstruct the hidden scene.Source: C. Plan-Conditioned Recurrent Student · p. 4

D. Learning from Privileged Supervision

  • Training simulates self-occlusion, token dropout, and scheduled multi-decision blackouts, exposing the recurrent policy to consecutive decisions without current object geometry.Unavailable object geometry is zeroed, and observation age tracks missing visibility.Source: D. Learning from Privileged Supervision · p. 4
  • Behavior cloning pairs student-available inputs from teacher-controlled trajectories with the frozen teacher's categorical action distributions.The initial dataset uses partial observations, fixed-rollout context, end-effector features, and the previous action.Source: D. Learning from Privileged Supervision · p. 5
  • DAgger lets the current student control execution and uses the frozen teacher to label student-visited states, extending privileged supervision beyond teacher-controlled states.The teacher output is a supervised target; the student's selected primitive remains the executed action.Source: D. Learning from Privileged Supervision · p. 5
  • Both stages minimize a recovery-weighted forward-KL objective, with valid-step masking that keeps occluded observations eligible for supervision.Matching the full teacher distribution preserves relative preferences among corrective actions beyond its argmax.Source: D. Learning from Privileged Supervision · p. 5, Figure 4 · p. 5

V. EXPERIMENTS

The experiments are organized around comparative performance, supervision value, policy architecture under partial observation, and transfer to physical retrieval.

V. EXPERIMENTS

  • The evaluation asks whether plan-conditioned closed-loop execution improves over nominal replay and existing retrieval methods.This is one of the four stated experimental questions.Source: §V · p. 5
  • The experiments test whether student-visited-state supervision adds benefit beyond additional expert data.The comparison concerns the value of supervision on states encountered by the student.Source: §V · p. 5
  • The evaluation examines recurrent memory and plan-context horizon under partial observation, and whether the advantages transfer to physical retrieval without repeated scene reacquisition.These are the remaining stated experimental questions.Source: §V · p. 5

A. Experimental Protocol

The evaluation uses 11 movable objects in a 0.448×0.448 m workspace, with separate training, validation, and test scenes. Test conditions include perturbed execution scenes and partial observations with occlusion, token dropout, and a blackout.

Environment and scenes

  • Experiments use 2,097 training, 178 validation, and 511 independently generated test scenes with 11 movable objects in a 0.448×0.448 m workspace.The 16 motion primitives are shown in Fig. 2; the test set is reserved for final evaluation.Source: A. Experimental Protocol · p. 5

Planning–execution mismatch

  • Each nominal rollout is generated from the planning scene, while execution objects are independently perturbed by up to ±15 mm per planar axis and ±10° in yaw.Paired policies receive identical perturbations.Source: A. Experimental Protocol · p. 5

Partial-observation condition

  • At test time, self-occlusion is combined with 10% token dropout and one five-decision blackout sampled across the 120-decision horizon.Visible objects use exact simulator geometry, isolating missing observations from pose-estimation error.Source: A. Experimental Protocol · p. 6

Training and evaluation

  • The teacher is trained with PPO and frozen; the student is initialized by behavior cloning and refined with DAgger, with model selection on the 178 validation scenes.Each student fit uses 10,000 optimizer updates with minibatches of 32 recurrent sequences.Source: A. Experimental Protocol · p. 6

B. Plan-Conditioned Execution and Reference Policies

The evaluation compares TRACE with replay of its nominal rollout, planning-based methods, target-centric heuristics, and a privileged closed-loop teacher. TRACE pairs the same nominal prediction with partial closed-loop observations.

Reference policies

  • The references include internal policies, planning-based baselines, target-centric heuristics, and the privileged ONLINE TEACHER.The comparison is framed around plan-conditioned closed-loop execution versus replay of the same nominal rollout.Source: B. Plan-Conditioned Execution and… · p. 6
  • The qualitative robot comparison shows TRACE succeeding without online arm retraction, while ONLINE TEACHER succeeds after repeated complete-scene reacquisition.TEACHER REPLAY ends ungraspable, Spiral exits the workspace, and PMBS reaches its 15-push limit.Source: Figure 5 · p. 6

Simulation comparison

  • 90.7% success for TRACE versus 43.4% for TEACHER REPLAY on the 511-scene test set; TRACE reaches 96.7% for ONLINE TEACHER and exceeds PMBS at 88.1%.TRACE uses the same nominal prediction as replay, together with partial closed-loop observations.Source: Table I · p. 6, B. Plan-Conditioned Execution and… · p. 6

C. Effect of Student-State Supervision

Matched-label comparisons distinguish student-state supervision from simply adding expert demonstrations. DAgger improves success at equal label and optimization budgets, and also outperforms expert-only cloning with a larger label budget.

Expert-only cloning control

  • Expert BC reaches 78.8% success with 111,555 labels, compared with 90.7% for DAgger R3.The matched-label control and this larger-budget comparison support student-state coverage as the main source of DAgger’s advantage.Source: C. Effect of Student-State Supervision · p. 6

DAgger rounds

  • Most of DAgger’s gain appears after the first aggregation round, with smaller improvements in later rounds.Each round adds teacher labels on states induced by the current student.Source: C. Effect of Student-State Supervision · p. 6

Matched-label comparison

D. Plan-Context Horizon and Recurrent Memory

TRACE retains much of the privileged teacher’s reliability on hardware while avoiding repeated complete-scene reacquisition during pushing. Ablations support recurrent memory with local rollout context; the evidence is limited to planar clutter with known footprints and an initially visible scene.

D. Plan-Context Horizon and Recurrent Memory

  • Removing the GRU primarily increases execution-budget exhaustion, while success is similar across tested plan horizons and the full rollout shows no observed benefit over K = 4.The study retains K = 4 as a compact local context; it does not treat a plan-free controller as a matched ablation.Source: D. Plan-Context Horizon and Recurrent… · p. 7

E. Real-Robot Evaluation

  • The hardware benchmark uses 20 unseen clutter arrangements with two trials per scene and method; success requires grasping and lifting the target while maintaining workspace constraints.The fixed camera supplies partial observations throughout execution, while methods requiring complete scene state retract the arm to restore visibility.Source: E. Real-Robot Evaluation · p. 7

VI. CONCLUSIONS

  • On hardware, the one-time rollout also supplies a scene-specific execution horizon without a learned stop action.A clean initial observation initializes the digital twin and generates the rollout before manipulation.Source: §VI · p. 7, E. Real-Robot Evaluation · p. 7

E. Real-Robot Evaluation

  • 90.0% hardware success for TRACE versus 95.0% for ONLINE TEACHER, with total execution times of 67.3 s and 192.7 s, respectively.TRACE uses no arm retractions for state reacquisition during pushing; ONLINE TEACHER averages 16.8 per trial.Source: E. Real-Robot Evaluation · p. 7, Table IV · p. 8

VI. CONCLUSIONS

  • The evidence is limited to planar clutter with known object footprints and an initially visible scene; uncertain initial estimates and unfamiliar geometries need further evaluation.The paper identifies selective reacquisition when a rollout becomes uninformative, continuous actions, learned termination, and non-planar clutter as future work.Source: §VI · p. 8
Loading 2609.38857v1…