PaperlayerView exact revision on arXiv

Source-linked AI summary

TILDE: Teacher-in-the-Loop Distillation from Embodied Foundation Models

Yuxiang Lu, Yantong Liu, Jinghua Hou, Xueli Sun, Ruihua Han, Zhe Liu, Hengshuang Zhao

arXiv:2610.25376v1cs.ROcs.LG

TL;DR

Large-scale action pretraining gives EFMs strong manipulation capabilities, but reproducing it requires substantial robot data and computation. TILDE combines student-state supervision with teacher-executed recovery, while AutoDAgger automates intervention; TILDE improves real-world performance and AutoDAgger speeds recovery collection.

  • Problem

    Transferring pretrained EFM capabilities to new policies is challenging because action pretraining depends on substantial robot data and computation, and gains do not consistently extend to long-horizon, perturbed, or real-world settings.

    Source: §1 · p. 2

  • Method

    TILDE combines demonstrations, teacher supervision at student-induced states, and teacher-executed recovery trajectories, with AutoDAgger monitoring task progress to trigger teacher intervention.

    Source: §1 · p. 2, §3.3 · p. 5

  • Results

    Across three real-world tasks, TILDE improves over behavior cloning by up to 4.1×, while AutoDAgger collects recovery data more than 2× faster than HG-DAgger.

    Source: §1 · p. 3, §5 · p. 10

  • Takeaways & Limitations

    Cross-architecture experiments with SmolVLA validate capability transfer to heterogeneous student policies.

    Source: §5 · p. 10

    The evaluation covers a limited set of real-world tasks, simulation benchmarks, and teacher–student architectures.

    Source: C LIMITATIONS AND FUTURE WORK · p. 22

Abstract

from arXiv · show

Large-scale action pretraining equips Embodied Foundation Models (EFMs) with strong manipulation capabilities, with particular strengths in generalization and closed-loop robustness, but reproducing such pretraining requires substantial robot data and computation. To transfer these capabilities from pretrained EFMs to policies without action pretraining, we introduce TILDE (Teacher-In-the-Loop Distillation from EFMs), which synergizes on-policy distillation at student-induced states with learning from teacher-executed recovery trajectories. We further propose AutoDAgger, an autonomous data aggregation pipeline in which a robotic reward model is utilized to monitor task progress and trigger teacher intervention, thereby collecting recovery experience without human monitoring or manual control. AutoDAgger accelerates recovery-trajectory collection by more than $2\times$ compared with the human-in-the-loop method. By unifying learning from demonstrations, on-policy teacher supervision, and teacher's corrective behavior, TILDE improves task performance by up to $4.1\times$ over behavior cloning across three real-world tasks and supports effective cross-architecture distillation.

1 INTRODUCTION

TILDE transfers pretrained EFM capabilities through demonstrations, student-induced rollouts, and teacher-executed recovery trajectories. AutoDAgger automates recovery collection, and experiments report improved real-world task performance and cross-architecture transfer.

1 INTRODUCTION

  • Pretrained EFMs offer manipulation and generalization capabilities, but reproducing action pretraining requires diverse robot data and substantial computation.The paper asks whether these capabilities can be transferred without repeating expensive pretraining.Source: §1 · p. 2
  • Fixed demonstration supervision can leave closed-loop students unprepared for out-of-distribution states caused by compounding errors and environmental stochasticity.Interactive methods query learner-induced states, but OPD does not execute teacher corrections and human-in-the-loop recovery collection requires continuous monitoring and teleoperation.Source: §1 · p. 2
  • TILDE transfers teacher knowledge across human demonstrations, student-induced rollouts, and teacher-generated recovery experience.Offline demonstrations and teacher recovery trajectories provide behavior-cloning and distillation targets, while student rollout states receive on-policy teacher supervision only.Source: §1 · p. 2
  • AutoDAgger uses reward-model task monitoring to trigger teacher takeover and collect recovery trajectories without expensive human oversight.Teacher takeover avoids the reaction and teleoperation overhead of human intervention.Source: §1 · p. 2
  • TILDE distills directly in the common action space, requiring no internal-representation alignment between heterogeneous teacher and student architectures.This supports the framework's cross-architecture capability transfer.Source: §1 · p. 3
  • Up to 4.1× over behavior cloning, TILDE improves student performance on real-world manipulation tasks and supports cross-architecture capability transfer.The introduction also reports consistent gains over offline and on-policy distillation.Source: §1 · p. 3

2 RELATED WORK

Related work spans embodied foundation models, policy distillation, and interactive imitation learning. TILDE builds on student-state supervision by adding teacher-executed recovery trajectories and autonomous intervention.

2 RELATED WORK

  • EFMs integrate visual perception, language understanding, and low-level control; VLA models map multimodal observations and instructions to actions, while WAMs jointly model future states and actions.Training foundation models from scratch requires substantial data and computational resources.Source: §2 · p. 3
  • Policy distillation transfers teacher predictions or representations to students, with methods using model-level alignment, teacher-generated supervision, or on-policy teacher queries at student-induced states.TILDE extends OPD by executing teacher takeover and collecting the subsequent recovery trajectory.Source: §2 · p. 3
  • Interactive imitation learning queries experts on learner-induced states to mitigate distribution shift, but existing approaches often rely on policy-specific techniques or human oversight.AutoDAgger instead detects intervention needs and executes corrective actions; its pipeline monitors progress and selects successful teacher recoveries for aggregation.Source: §2 · p. 3, Figure 2 · p. 4

3 METHODOLOGY

TILDE combines offline demonstrations, student-induced states, and teacher-executed recovery trajectories, with AutoDAgger monitoring execution and triggering teacher intervention. It distills teacher actions directly in the shared action space, using supervision matched to each sample source.

3.1 PRELIMINARIES

  • Policies map RGB images, robot proprioception, and language instructions to action chunks that are repeatedly updated from new observations during closed-loop execution.Source: §3.1 · p. 4

3.2 OVERVIEW

  • TILDE first collects student rollouts and teacher recovery trajectories, then trains the student on them alongside existing demonstrations.The teacher is finetuned for the embodiment and task, then frozen; the student receives no robot action pretraining.Source: §3.2 · p. 4

3.3 AUTODAGGER

  • AutoDAgger uses a robotic reward model to monitor task progress and success, triggering teacher takeover after sustained stagnation and retaining successfully recovered episodes.Progress counts as meaningful only when it exceeds the reference by a margin; episodes without confirmed teacher recovery are discarded.Source: §3.3 · p. 5

3.4 DISTILLATION OBJECTIVE

  • TILDE trains from demonstrations, student-rollout states, and teacher recovery trajectories, applying the student’s original action-generation objective to teacher-predicted actions.Distillation is imposed on action outputs rather than internal representations.Source: §3.4 · pp. 5–6
  • TILDE outperforms all student baselines on all three real-world tasks, with the teacher policy serving as a reference.Source: Figure 3 · p. 6
  • Demonstration and recovery samples combine behavior cloning from executed actions with teacher distillation, while student-rollout states receive only teacher distillation.Source: §3.4 · p. 6

4 EXPERIMENTS

Experiments evaluate TILDE on three real-world manipulation tasks, where it outperforms the evaluated baselines; analyses examine recovery behavior, data scaling and quality, automated collection, and cross-architecture transfer.

4.1 EXPERIMENTAL SETUP

  • Experiments use three manipulation tasks, fixed test cases, and a π0.5 teacher; unless otherwise specified, the student has the same architecture but no action pretraining.The task monitor uses Robometer without task-specific finetuning.Source: §4.1 · p. 6

4.2 MAIN RESULTS

  • 63.9, 80.6, and 69.2 are TILDE’s scores on Fold Dishcloth, Pick Pen, and Pick Beverage, respectively, approximately 3.07×, 2.18×, and 4.14× the BC baseline scores.TILDE outperforms all evaluated baselines on each task.Source: §4.2 · p. 7

4.3 ANALYSIS

  • Adding student-rollout supervision or teacher recovery trajectories improves performance across all three tasks, and combining them yields the highest scores.Recovery BC adds imitation of recorded recovery actions; Recovery Distill also distills on demonstrations and recovery samples, further improving performance.Source: §4.3 · p. 7, Figure 4 · p. 7
  • TILDE succeeds from 8 of 12 predefined Fold Dishcloth error states, versus 4/12 for BC, 5/12 for offline distillation and OPD, and 9/12 for the pretrained teacher.The evaluation tests recovery from states that deviate from nominal execution.Source: §4.3 · p. 8
  • On Fold Dishcloth, the score rises from 27.8 with offline distillation to 63.9, 65.3, and 77.8 after aggregating 30, 60, and 90 AutoDAgger episodes.At 90 episodes, the student narrows the gap to the teacher score of 87.5.Source: Table 2 · p. 7, §4.3 · p. 8
  • With 30 additional episodes per task, AutoDAgger data yield higher BC and distillation scores than additional human demonstrations across all three tasks.With distillation, AutoDAgger scores are 63.9, 80.6, and 69.2 on Fold Dishcloth, Pick Pen, and Pick Beverage.Source: §4.3 · p. 8, Table 3 · p. 8
  • AutoDAgger reduces recovery-trajectory collection time versus HG-DAgger while achieving comparable recovery success rates and automating monitoring and takeover.On Fold Dishcloth, 60 AutoDAgger episodes reach a score of 65.3, matching HG-DAgger with 30 episodes; recovery collection is 2.53× faster.Source: Table 4 · p. 8, §4.3 · p. 9
  • With SmolVLA as student, TILDE improves LIBERO-10 success from 64.6% to 79.6%; Robometer also predicts rollout outcomes across four LIBERO suites, with recall decreasing on LIBERO-10.The monitor evaluation uses 200 SmolVLA rollout episodes per suite and supports policy-independent monitoring without task-specific finetuning.Source: §4.3 · p. 9, Table 5 · p. 9, Table 6 · p. 9

5 CONCLUSION

TILDE transfers pretrained EFM capabilities through student-induced supervision and teacher-executed recovery trajectories. Across three real-world tasks it improves over several distillation baselines, while AutoDAgger collects recovery data faster than human-in-the-loop collection.

5 CONCLUSION

  • TILDE combines supervision at student-induced states with teacher-executed recovery trajectories, covering corrective actions and subsequent recovery-state distributions.It targets transferring pretrained EFM capabilities to student policies without repeating foundation-scale pretraining.Source: §5 · p. 9
  • Across three real-world manipulation tasks, TILDE improves over behavior cloning, offline distillation, and conventional on-policy distillation.Cross-architecture experiments with SmolVLA also validate transfer to heterogeneous student policies.Source: §5 · p. 10
  • AutoDAgger collects recovery data more than 2× faster than HG-DAgger without human monitoring or takeover.Task monitoring and execution management support autonomous collection.Source: §5 · p. 10

APPENDIX

  • The appendix covers AutoDAgger and experiment implementation, additional results, and limitations and future directions.Its sections address implementation and benchmarks, failure-detection evaluation and collection visualizations, and limitations with future work.Source: APPENDIX · p. 15

A ADDITIONAL IMPLEMENTATION DETAILS

AutoDAgger monitors policy execution using bounded visual histories and task-progress predictions. Its execution manager detects stagnation, manages teacher takeover and recovery, and retains successful recovery trajectories for training.

A.1 AUTODAGGER

  • Each monitor update evaluates a bounded snapshot of visual history, submitted at least τ_mon seconds apart, while asynchronous evaluation does not block policy inference.When the history exceeds N_frame observations, it is downsampled to N_frame evenly spaced observations; experiments fix N_frame = 8 and τ_mon = 1 s.Source: §A.1 · p. 15
  • The execution manager continuously checks stagnation and can trigger intervention as soon as the failure timeout is reached, without waiting for another monitor update.Failure detection begins after episode initialization; the progress reference and valid-execution timestamp are then updated for subsequent monitor outputs.Source: §A.1 · p. 15
  • On student failure, AutoDAgger stops the robot, clears pending student actions, transfers control to the teacher, and restarts monitoring from a fresh state.Execution resumes only when a teacher action chunk is available, preventing residual student actions after takeover.Source: §A.1 · p. 15
  • Successful episodes are retained as combined student-rollout and teacher-recovery trajectories; failed teacher recoveries and student-only successes are discarded.AutoDAgger specifically targets recovery-data collection.Source: §A.1 · p. 15
  • Success is confirmed only after predictions remain above the success threshold for a continuous duration, reducing false detections from transient spikes.A prediction at or below the threshold resets the timer, and success is evaluated upon a new monitor output.Source: §A.1 · p. 16
  • Task- and policy-specific failure and success timeouts accommodate different execution dynamics, with shorter student failure timeouts enabling earlier intervention.Teacher timeouts are longer to allow recovery; task thresholds account for differences such as dishcloth pauses versus more compact picking progress.Source: §A.1 · p. 16, Table 7 · p. 16

A.2 REAL-WORLD EXPERIMENTS

The real-world experiments evaluate three manipulation tasks on a dual-arm PiperX platform with a fixed rollout protocol. They measure both complete-task success and a staged score that captures partial completion and execution quality.

A.2 REAL-WORLD EXPERIMENTS

  • The platform uses four 7-DoF PiperX arms, one front-facing D455 camera, and two wrist-mounted D435 cameras.Two leader arms are human-operated; the corresponding follower arms reproduce commands during collection and execute policy actions during inference, as illustrated in Figure 5.Source: §A.2 · p. 16, Figure 5 · p. 17
  • Real-robot evaluation covers Fold Dishcloth, Pick Pen, and Pick Beverage, spanning deformable-object manipulation, precision grasping, and object relocation.The task illustrations are shown in Figure 6.Source: §A.2 · pp. 16–17, Figure 6 · p. 18
  • Each method receives 12 Fold Dishcloth, 27 Pick Pen, and 30 Pick Beverage rollouts, with success requiring full task completion within the episode limit and no human intervention.Test configurations use fixed object categories and controlled position and orientation variations, kept as consistent as practically possible across methods.Source: §A.2 · p. 17
  • The fine-grained task score averages stage scores across stages and rollouts, then multiplies by 100; each stage is scored 0, 0.5, or 1.Scores represent failure, partial completion or completion after repeated attempts, and single-attempt success including immediate recovery from minor issues.Source: §A.2 · p. 17
  • Fold Dishcloth is scored across two folding stages and a final pickup-and-placement stage, distinguishing repeated attempts and partial recovery from single-attempt completion.The final stage places the folded cloth in the front-left target region; the scoring examples include recovery from a flipped corner.Source: §A.2 · p. 17
  • Pick Pen and Pick Beverage each use separate grasping and placement stages, with intermediate scores for recovery attempts or undesirable contact.Pens are placed in a yellow cup and beverages in a paper box; multiple-pen trials apply the two-stage procedure independently to each pen.Source: §A.2 · pp. 17–18

A.3 SIMULATION BENCHMARKS

The appendix describes LIBERO and LIBERO-Plus evaluations, including joint training across four LIBERO suites and perturbation-based tests of out-of-distribution generalization.

A.3 SIMULATION BENCHMARKS

  • A single policy is trained jointly across LIBERO’s Spatial, Object, Goal, and 10 (Long) suites, each comprising 10 tasks evaluated over 50 rollouts.Dataset sources differ by policy: π0.5, VLA-Adapter-Pro, and SmolVLA use the OpenVLA-released dataset, while X-VLA uses its authors’ HDF5 dataset.Source: §A.3 · p. 19
  • LIBERO-Plus evaluations use Background scene-appearance perturbations and Robot initial-pose perturbations on checkpoints trained only with original LIBERO data.The settings contain 1,076 and 1,550 task instances, respectively, and test out-of-distribution generalization under environmental and embodiment perturbations.Source: §A.3 · p. 19

A.4 TRAINING SETUP

The setup specifies model initialization, source-dependent supervision, training configurations, and deployment conditions for the experiments.

A.4 TRAINING SETUP

  • Action-pretraining-free π0.5 and X-VLA variants start from PaliGemma and Florence-2-large VLM backbones, respectively, with action experts trained from scratch.The models are trained from official codebases with modifications for the pipeline and data formats.Source: §A.4 · p. 19
  • Training follows model-specific recommended configurations, samples pooled datasets randomly without balancing, and sets λKD = 1 across experiments.Real-robot demonstrations are recorded in absolute joint space at 30FPS and converted to relative joint actions when needed.Source: §A.4 · p. 19
  • AutoDAgger collection uses a fixed behavior-cloning student trained on demonstrations, followed by reinitialization and training on the aggregated dataset.The initial collection policy is trained on Ddemo using behavior cloning only.Source: §A.4 · p. 19
  • Distillation matches shared-noise flow-matching velocities for π0.5 teacher–student pairs, but uses denoised teacher actions for cross-architecture students.Cross-architecture students use the imitation-learning objective applied to ground-truth actions.Source: §A.4 · p. 19

A.5 ROBOT DEPLOYMENT SETUP

  • Training uses eight NVIDIA A800 GPUs; deployment uses an RTX 4090 workstation and a 30Hz ROS client with asynchronous inference and temporal ensembling.Reported latency is 80.0±1.6ms for π0.5 and 266.9±0.5ms for Robometer.Source: §A.4 · p. 19, §A.5 · p. 19

A.6 CROSS-ARCHITECTURE DISTILLATION SETUP

The cross-architecture experiment uses π0.5 as teacher and SmolVLA as student, applying the collection and training protocol to LIBERO.

A.6 CROSS-ARCHITECTURE DISTILLATION SETUP

  • The LIBERO-trained SmolVLA behavior-cloning student collects AutoDAgger rollouts under a π0.5 teacher, then TILDE trains on demonstrations, student rollouts, and teacher recoveries.The collection and training protocol follows the preceding setup.Source: §A.6 · p. 20

B.1 VALIDATION OF TASK MONITOR ON REAL-WORLD ROLLOUTS

On 20 real-robot rollouts, the task monitor achieved 85% accuracy, with conservative success confirmation; representative collection trajectories show takeover after stalled progress and low success probability.

B.1 VALIDATION OF TASK MONITOR ON REAL-WORLD ROLLOUTS

  • The monitor evaluation comprises 20 rollouts: 10 student and 10 teacher, with two student successes and eight teacher successes.Ground-truth outcomes were manually determined from task completion.Source: §B.1 · p. 20
  • 85% overall accuracy: the monitor identified 8/10 successful and 9/10 failed rollouts, with 88.9% precision, 80.0% recall, and 84.2% F1.The evaluation set contains 10 successful and 10 failed trajectories.Source: §B.1 · p. 21, Table 11 · p. 21
  • Two false negatives reflect conservative success confirmation, which reduces lower-quality recovery data entering training but lowers success recall.The authors note the evaluation set is limited by the cost of real-robot experiments.Source: §B.1 · p. 21

B.2 EXPERIMENTAL RESULT STATISTICS

  • Tables 12–15 supplement the main-text experiments with performance and recovery-data collection statistics.Source: §B.2 · p. 21

B.3 VISUALIZATION OF AUTODAGGER COLLECTION

  • Across Fold Dishcloth, Pick Pen, and Pick Beverage examples, takeover follows stalled progress and low success probability; success probability rises above θs near completion.Reported recovery progress reaches 0.869, 0.814, and 0.847; success probabilities reach 0.844, 0.582, and 0.926, respectively.Source: §B.3 · p. 21

C LIMITATIONS AND FUTURE WORK

The authors identify limits in AutoDAgger’s collection regime, its dependence on monitoring and teacher recoverability, and the breadth of evaluation to date.

C LIMITATIONS AND FUTURE WORK

  • TILDE is evaluated with one AutoDAgger collection round rather than iterative aggregation, leaving repeated policy-update and collection cycles for future work.The authors suggest additional rounds could expose the student to increasingly relevant failure and recovery states.Source: C LIMITATIONS AND FUTURE WORK · p. 21
  • AutoDAgger depends on task-monitor accuracy and teacher recoverability; incorrect interventions or unrecoverable states may reduce collected-data quality or efficiency.More robust monitoring and adaptive intervention are identified as future directions.Source: C LIMITATIONS AND FUTURE WORK · p. 22
  • Evaluation covers a limited set of real-world tasks, simulation benchmarks, and teacher–student architectures.The authors propose scaling evaluation to broader tasks, embodiments, and heterogeneous policy families.Source: C LIMITATIONS AND FUTURE WORK · p. 22
Loading 2610.25376v1…