MoE-DAgger: Transition-Aware Router-Entropy Gating for Selective Human Correction in Phase-Structured Manipulation

Sungjun Heo, Sungjun Yang, Heedon Jeong, Junhyoung Ha, Jeong hwan Jeon

November 2026

Green task rollout on the lower shelf, external and wrist views.

TL;DR MoE-DAgger decides when to ask a human for a correction by tracking how the router of a phase-supervised mixture-of-experts policy moves between task phases, so normal phase handoffs do not trigger queries. With matched human-control time it reaches 84% and 78% success on two real Franka pushing tasks, 22 and 26 points above entropy gating without phase awareness.

What they do

In DAgger with human takeover, the costly part is the operator’s time, so the robot should ask for help only on rollouts that are going wrong, and early enough to matter. The authors use a policy with one action head per task phase and a soft router that blends them. Router entropy alone is a poor trigger: the same value can come from a normal handoff between consecutive phases or from the router being split between unrelated phases. MoE-DAgger uses the known phase order to tell these apart.

How it works

Policy. A shared encoder feeds five phase-anchored ACT heads (approach, entry/alignment, push, withdrawal/clearance, home return); the router’s soft weights blend their action chunks. The policy never receives a phase label or command at run time.
Gate. From causal router history only, a tracker follows the current phase and its legal successor. A transition between the two gets a bounded time to settle and raises no query.
Triggers. A correction is requested when averaged entropy stays high outside such a handoff, when a handoff times out, or when routing persistently favors an earlier phase.
Correction. The robot holds, the human controls the rest of the rollout, and the policy is retrained on the original data plus the corrections.

Results

Franka Research 3, two shelf-pushing tasks (Green / Red), 30 demonstrations each plus 112.55 s / 127.55 s of corrections, 50 rollouts per method and task: 84.0% / 78.0% task success with no collisions, against 62.0% / 52.0% success and 14% / 16% collisions for phase-unaware entropy.
ACT temporal disagreement is close at 82.0% / 72.0%, also with no collisions; the confidence intervals of the difference span zero.
On 50 recorded traces per task, 83.3% (15 selected) and 80.0% (25 selected) of first proposals land on failed rollouts inside the failure window, against 70.0% / 72.0% for ACT disagreement and 63.3% / 60.0% for ensemble disagreement.
Removing only the handoff veto drops success to 68.0% / 60.0%.

Limitations

One robot and fixture, separate per-task policies, and two geometric variants of one pushing task with a known linear phase order. Each method has a single acquired dataset and one aggregation round. The budget excludes offline phase annotation. Partial-push failures are poorly covered (9/20 on Red), and the critical-window analysis does not show that an intervention would have recovered the rollout.

Videos

Red task rollout on the upper shelf, external and wrist views.

Green task rollout with the router phase probabilities and entropy plotted live.

Figures

From Figure 1: Equal router entropy can require different acquisition decisions. Green: The robot remains autonomous ... — From Figure 1: Equal router entropy can require different acquisition decisions. Green: The robot remains autonomous while router support shifts from P0 to P1 and settles within the allowed transition time. Red: The same entropy value can arise when the router supports the nonadjacent phases P0 and P4. The acquisition rule does not issue a query during the settling adjacent transition, but requests human correction if the nonadjacent support persists.

From Figure 2: Online MoE-DAgger acquisition architecture. Separate policies for Green and Red map synchronized four-... — From Figure 2: Online MoE-DAgger acquisition architecture. Separate policies for Green and Red map synchronized four-view RGB history and proprioception to a soft mixture of outputs from five phase-anchored action heads. The acquisition module receives only causal router history. It permits a bounded, settling transition between the current phase and its legal successor, but requests human correction when high entropy persists outside that case, when the transition times out, or when routing repeatedly moves backward.

From Figure 3: Ordered phase structure of the two real-robot pushing tasks. Top: Green operates on the lower shelf; b... — From Figure 3: Ordered phase structure of the two real-robot pushing tasks. Top: Green operates on the lower shelf; bottom: Red operates on the upper shelf. Both tasks follow the semantic sequence P0 approach, P1 entry/alignment, P2 push, P3 withdrawal/clearance, and P4 home return. Representative external and wrist-camera keyframes are shown in execution order.

From Figure 4: Failure coverage and successful-rollout burden on the same 50 traces per task, against the number of s... — From Figure 4: Failure coverage and successful-rollout burden on the same 50 traces per task, against the number of selected rollouts B. Panels (a,b) count aligned failures, panels (c,d) count successful rollouts receiving a proposal, and panels (e,f) count aligned collision/no-contact failures. Higher is better in the top and bottom rows and lower in the middle row.

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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong