ChunkTrust

ChunkTrust

We treat action-chunk execution horizon as a latent decision and select executable prefixes using the policy's own denoising dynamics, boundary continuity, and a phase-aware posterior.

Core Idea Evidence Method Results Real Robot Case Study Implementation Citation

Core Idea

Execution horizon is a decision, not a constant.

Fixed action chunks can over-commit in contact-rich phases and replan too often during stable motion. ChunkTrust estimates how much of each predicted chunk can be trusted before the next closed-loop correction.

+9.5 Success Rate on RoboTwin2.0 pi0.5, 29.6 to 39.1
+9.7 Success Rate on RoboCasa Qwen3GR00T, 47.8 to 57.5
+7.1 Process Score on real robots, 50.4 to 57.5
ChunkTrust teaser figure showing adaptive horizon behavior.
ChunkTrust adapts the number of actions committed from each chunk using action-expert evidence, shortening execution when feedback is needed and extending it when the predicted prefix is stable.

Action-Expert Evidence

Two complementary signals decide how far to execute.

01

Intra-chunk spectral stability

Fourier analysis of the denoising velocity trace estimates whether a candidate prefix remains smooth inside the generated chunk.

02

Inter-chunk continuity

Speed uniformity across the executed-history and predicted-prefix boundary penalizes horizon choices that introduce abrupt execution changes.

03

Phase-aware posterior

A kernel-forgotten Beta posterior tracks which horizons are reliable in the current manipulation phase instead of averaging over the full episode.

Rationality Check

The evidence tracks when longer chunks are trustworthy.

Metric-level analysis supports the intuition that reliable horizons vary by phase: short horizons are preferred around unstable contacts, while longer horizons become useful in smooth transport or post-contact motion.

Metric rationality figure for fixed K equals 50 episode-level analysis.

Method

AHS scores candidate horizons online. QHA amortizes the preference.

ChunkTrust evaluates how far each predicted action chunk should be trusted before replanning. The training-free Action-aware Horizon Selector scores candidate prefixes from intra-chunk spectral stability and inter-chunk continuity, then tracks phase-dependent reliability with an online posterior. The Query-based Horizon Adapter distills these preferences into a lightweight horizon prior that can be fused with fresh online evidence at deployment time.

Overview diagram of AHS and QHA horizon adaptation.

Results

AHS improves every evaluated simulated base policy.

Full per-task success rates are reported below for RoboTwin2.0 and RoboCasa GR1 Tabletop. Yellow cells indicate AHS, and blue cells indicate AHS+QHA.

Per-task success rates on RoboTwin2.0

Task pi0 pi0 AHS pi0 AHS+QHA pi0.5 pi0.5 AHS pi0.5 AHS+QHA Fast-WAM Fast-WAM AHS
EasyHard EasyHard EasyHard EasyHard EasyHard EasyHard EasyHard EasyHard
Blocks Ranking RGB 190 28 +91 +1 31 +125 +5 3620 48 +1229 +9 56 +2033 +13 9998 100 +1100 +2
Handover Block 4110 41 +05 -5 64 +239 -1 4414 44 +014 +0 32 -1210 -4 9481 91 -382 +1
Handover Mic 1002 100 +024 +22 100 +030 +28 9864 100 +256 -8 98 +059 -5 10099 100 +0100 +1
Hanging Mug 173 14 -39 +6 23 +610 +7 148 13 -112 +4 13 -114 +6 6664 71 +569 +5
Place A2B Left 241 34 +100 -1 39 +151 +0 414 47 +64 +0 47 +615 +11 9595 93 -293 -2
Place Bread Basket 118 18 +716 +8 15 +410 +2 3017 50 +2029 +12 55 +2541 +24 9191 93 +294 +3
Place Bread Skillet 152 23 +85 +3 23 +82 +0 2410 36 +1219 +9 44 +2019 +9 9191 94 +394 +3
Place Can Basket 335 22 -112 -3 33 +03 -2 3515 50 +1529 +14 55.2 +20.234 +19 7163 69 -267 +4
Average 32.53.9 35.0 +2.57.8 +3.9 41.0 +8.58.8 +4.9 40.219.0 48.5 +8.324.0 +5.0 50.0 +9.828.1 +9.1 88.485.2 88.9 +0.587.4 +2.1
Overall 18.2 21.4 +3.2 24.9 +6.7 29.6 36.2 +6.6 39.1 +9.5 86.8 88.1 +1.3

Full per-task success rates on RoboCasa GR1 Tabletop

Task QwenFAST + Qwen3VL QwenPI + Qwen3VL N1.5 N1.5 AHS N1.6 N1.6 AHS Qwen3GR00T + Qwen3VL Qwen3GR00T AHS
PnPBottleToCabinetClose38.026.064.068.0 +4.051.554.0 +2.546.064.0 +18.0
PnPCanToDrawerClose44.062.018.012.0 -6.013.012.0 -1.080.080.0 +0.0
PnPCupToDrawerClose56.042.012.04.0 -8.08.514.0 +5.554.052.0 -2.0
PnPMilkToMicrowaveClose44.050.038.034.0 -4.014.020.0 +6.048.042.0 -6.0
PnPPotatoToMicrowaveClose14.042.054.036.0 -18.041.550.0 +8.528.028.0 +0.0
PnPWineToCabinetClose14.032.016.020.0 +4.016.524.0 +7.546.052.0 +6.0
PnPNovelFromCuttingboardToBasket54.040.050.052.0 +2.058.054.0 -4.048.070.0 +22.0
PnPNovelFromCuttingboardToCardboardbox42.046.036.034.0 -2.046.546.0 -0.540.054.0 +14.0
PnPNovelFromCuttingboardToPan58.060.068.064.0 -4.068.580.0 +11.568.080.0 +12.0
PnPNovelFromCuttingboardToPot58.040.034.056.0 +22.065.064.0 -1.052.076.0 +24.0
PnPNovelFromCuttingboardToTieredbasket40.044.046.032.0 -14.046.554.0 +7.556.044.0 -12.0
PnPNovelFromPlacematToBasket36.044.050.046.0 -4.058.548.0 -10.542.054.0 +12.0
PnPNovelFromPlacematToBowl38.052.050.062.0 +12.057.560.0 +2.544.066.0 +22.0
PnPNovelFromPlacematToPlate42.050.062.066.0 +4.063.082.0 +19.048.072.0 +24.0
PnPNovelFromPlacematToTieredshelf18.028.014.026.0 +12.028.536.0 +7.518.020.0 +2.0
PnPNovelFromPlateToBowl52.052.058.058.0 +0.057.058.0 +1.060.060.0 +0.0
PnPNovelFromPlateToCardboardbox30.040.040.048.0 +8.043.558.0 +14.550.054.0 +4.0
PnPNovelFromPlateToPan48.036.044.048.0 +4.051.068.0 +17.054.054.0 +0.0
PnPNovelFromPlateToPlate50.048.066.074.0 +8.078.782.0 +3.370.074.0 +4.0
PnPNovelFromTrayToCardboardbox28.034.044.052.0 +8.051.548.0 -3.538.056.0 +18.0
PnPNovelFromTrayToPlate34.064.050.060.0 +10.071.068.0 -3.056.062.0 +6.0
PnPNovelFromTrayToPot46.044.046.050.0 +4.064.564.0 -0.550.066.0 +16.0
PnPNovelFromTrayToTieredbasket36.050.044.038.0 -6.057.056.0 -1.036.056.0 +20.0
PnPNovelFromTrayToTieredshelf16.028.034.038.0 +4.031.534.0 +2.516.044.0 +28.0
Average39.043.943.344.9 +1.747.651.4 +3.847.857.5 +9.7

Real Robot

AgileX COBOT Magic rollouts under clean and randomized settings.

We use pi0.5 as the policy, trained on clean real-robot demonstrations.

Real robot task setup on AgileX COBOT Magic.
Real robot performance chart comparing fixed horizon and AHS.

Fold Towels

Clean

Fold Towels

Randomized

Fold Towels

More Randomized

Bread to Plate

Clean

Bread to Plate

Randomized

Bread to Plate

More Randomized

Drink to Basket

Clean

Drink to Basket

Randomized

Drink to Basket

More Randomized

Duck to Drawer

Clean

Duck to Drawer

Randomized

Duck to Drawer

More Randomized

Case Study

Adaptive horizons follow manipulation phases.

AHS selects shorter horizons around grasping and transport corrections, then commits longer prefixes once motion becomes stable after contact.

Task prompt: use the left arm to grasp the red block on the table, handover it to the right arm and place it on the blue pad

Phase 1: approaching the red block.
Phase 1 The policy prepares to grasp the red block with the left arm
Phase 2: moving the block toward handover.
Phase 2 The policy moves the block toward the handover pose
Phase 3: stabilizing before handover.
Phase 3 The policy stabilizes the block before handover
Phase 4: regrasping with the right arm.
Phase 4 The policy regrasps the block with the right arm
Phase 5: carrying the block toward the target pad.
Phase 5 The policy carries the block toward the target pad
Phase 6: aligning the block for placement.
Phase 6 The policy aligns the block for placement on the blue pad
Phase 1 Step 0

The policy prepares to grasp the red block with the left arm

Selected horizon K=41
0 309
AgileX COBOT Magic real robot platform.

Implementation

Drop-in horizon selection for existing evaluation loops.

AHS runs as an evaluation-time wrapper around existing policy rollouts. It observes the policy's action trace and recent execution history, scores candidate prefixes, and returns the next horizon without changing the policy weights.

Citation

BibTeX

@misc{chunktrust2026,
  title  = {ChunkTrust},
  author = {Anonymous Authors},
  year   = {2026}
}