MoE-DAgger: Transition-Aware Router-Entropy Gating for Selective Human Correction in Phase-Structured Manipulation
Organizations: Artificial Intelligence Graduate School, Ulsan National Institute of Science and Technology (UNIST), Republic of Korea. · Department of Mechanical Engineering, Ulsan National Institute of Science and Technology (UNIST), Republic of Korea. · Department of Electrical Engineering, Ulsan National Institute of Science and Technology (UNIST), Republic of Korea.
Abstract
Selective data acquisition in Dataset Aggregation (DAgger) on robots requires deciding which learner rollouts to correct and when to intervene. In manipulation with ordered phases, router entropy from a phase-supervised soft mixture-of-experts (MoE) is insufficient: the same value can reflect either a normal adjacent-phase handoff or support for nonadjacent or earlier phases. We propose MoE-DAgger, which uses causal router history to distinguish these cases without providing the policy with a phase label, progress variable, or hard phase command during execution. The acquisition rule allows an adjacent transition to settle, but requests correction when high entropy persists outside such a transition, the transition times out, or routing persistently favors an earlier phase. Under matched active human-control budgets, MoE-DAgger achieves 84.0% task success on Green and 78.0% on Red, exceeding phase-unaware filtered entropy by 22 percentage points on Green and 26 points on Red, with no observed collisions in 50 rollouts per task. On 50 shared intervention-free traces per task, 83.3% and 80.0% of its first proposals fall on failed rollouts within predefined failure windows when 15 and 25 traces are selected, compared with 70.0% and 72.0% for temporal action-chunk disagreement and 63.3% and 60.0% for ensemble disagreement. Shadow replay and an active no-veto ablation indicate that suppressing queries at normal adjacent handoffs is a key contributor to the improved allocation.
Figures & tables
| Quantity | Value | Quantity | Value |
|---|---|---|---|
| Views / experts | 4 / 5 | Chunk / rate | 30 steps / 20 Hz |
| (G/R), | 0.50/0.46, 5 | 5/5 steps | |
| 0.15, 0.10 | 0.60, 0.35 | ||
| 3 steps, 0.05 | (G/R) | 4 steps, 30/30 steps | |
| (G/R) | 0.78/0.47 | (G/R) | 1.11/0.91 |
| 3 steps | jump / action query | disabled / disabled |
| Human-control (s) | Seq. execution (%) | Task success (%) | Signed margin (cm) | Collision (%) | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Method | G | R | G | R | G | R | G | R | G | R |
| BC-only | 0 | 0 | 88.0 [-0.35ex] [76.2,94.4] | 82.0 [-0.35ex] [69.2,90.2] | 52.0 [-0.35ex] [38.5,65.2] | 48.0 [-0.35ex] [34.8,61.5] | [-0.35ex] | [-0.35ex] | 12.0 [-0.35ex] [5.6,23.8] | 18.0 [-0.35ex] [9.8,30.8] |
| ExtraDemo-BC | 114.85 | 128.50 | 90.0 [-0.35ex] [78.6,95.7] | 82.0 [-0.35ex] [69.2,90.2] | 58.0 [-0.35ex] [44.2,70.6] | 46.0 [-0.35ex] [33.0,59.6] | [-0.35ex] | [-0.35ex] | 10.0 [-0.35ex] [4.3,21.4] | 18.0 [-0.35ex] [9.8,30.8] |
| Human-trigger DAgger | 114.45 | 127.95 | 98.0 [-0.35ex] [89.5,99.6] | 92.0 [-0.35ex] [81.2,96.8] | 72.0 [-0.35ex] [58.3,82.5] | 68.0 [-0.35ex] [54.2,79.2] | [-0.35ex] | [-0.35ex] | 0.0 [-0.35ex] [0.0,7.1] | 4.0 [-0.35ex] [1.1,13.5] |
| Phase-matched Random | 116.20 | 126.45 | 92.0 [-0.35ex] [81.2,96.8] | 90.0 [-0.35ex] [78.6,95.7] | 62.0 [-0.35ex] [48.2,74.1] | 60.0 [-0.35ex] [46.2,72.4] | [-0.35ex] | [-0.35ex] | 8.0 [-0.35ex] [3.2,18.8] | 10.0 [-0.35ex] [4.3,21.4] |
| Phase-unaware filtered- | 114.85 | 128.35 | 86.0 [-0.35ex] [73.8,93.0] | 84.0 [-0.35ex] [71.5,91.7] | 62.0 [-0.35ex] [48.2,74.1] | 52.0 [-0.35ex] [38.5,65.2] | [-0.35ex] | [-0.35ex] | 14.0 [-0.35ex] [7.0,26.2] | 16.0 [-0.35ex] [8.3,28.5] |
| Allocation | Mode coverage | ||||
| Rule | Align. | Succ. | Coll. | No-cont. | Partial |
| Green, B=15, (30% selected fraction) | |||||
| ACT temporal disagr. | 11/15 (73.3%) | 2/21 (9.5%) | 2/3 | 6/11 | 3/15 |
| Ensemble disagr. | 10/15 (66.7%) | 3/21 (14.3%) | 2/3 | 5/11 | 3/15 |
| Phase-unaware | 0/15 (0.0%) | 7/21 (33.3%) | 0/3 | 0/11 | 0/15 |
| MoE-DAgger | 12/15 (80.0%) | 1/21 (4.8%) | 2/3 | 7/11 | 3/15 |
| Variant | Edges | Prop. | Aligned fail/prop. | Handoff ovlp. |
|---|---|---|---|---|
| A. Green | ||||
| Raw | 314 | 50 / 50 (100.0%) | 0 / 50 (0.0%) | 147 / 147 (100.0%) |
| Avg. persist. | 170 | 50 / 50 (100.0%) | 0 / 50 (0.0%) | 97 / 147 (66.0%) |
| Complete w/o veto | 170 | 50 / 50 (100.0%) | 0 / 50 (0.0%) | 97 / 147 (66.0%) |
| Complete rule (shadow) | 36 | 32 / 50 (64.0%) | 24 / 32 (75.0%) | 0 / 147 (0.0%) |
| B. Red | ||||
| Acquisition gate | Human-control (s) | Veto q. /rollout (%) | Seq. execution (%) | Task success (%) | Collision (%) |
|---|---|---|---|---|---|
| No adjacent veto | 114.70 / 130.20 | 92.0 / 50.0 | 90.0 / 84.0 | 68.0 / 60.0 | 10.0 / 16.0 |
| Full MoE-DAgger | 112.55 / 127.55 | 0.0 / 0.0 | 100.0 / 98.0 | 84.0 / 78.0 | 0.0 / 0.0 |
| (Full – No-veto) |
Data-Efficient Robot Learning
Data-efficient robot learning develops ways for robots to acquire reliable skills with fewer demonstrations, interactions, and human annotations.
See all 205 papers