Recoleta
GitHub
Source note

Drift is a Sampling Error: SNR-Aware Power Distributions for Long-Horizon Robotic Planning

Authors
Kewei Chen, Yayu Long, Mingsheng Shang
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot PlanningInference Time ScalingMcmc SamplingLong Horizon Control
Open arXivMarkdown source

Summary

CAPS is a training-free inference-time method for VLA robot policies that reduces long-horizon instruction drift by searching over future action chunks when model uncertainty rises. It reports higher success rates than π0, π0.5, TACO, and several VLA baselines on RoboTwin and Simpler-WindowX.

Problem

  • Long-horizon VLA policies can choose locally likely actions that lose the original task goal after many steps.
  • This matters because one wrong action in bimanual or sequential manipulation can make the task unrecoverable.
  • Existing prompt and reranking methods depend on the sampled candidate set and do not iteratively repair a single trajectory.

Approach

  • CAPS treats drift as a sampling error in the policy's trajectory distribution.
  • It samples from a power distribution, π(τ) ∝ pθ(τ|I,Ht)^α, which gives more weight to trajectories the base model scores as consistent with the instruction and history.
  • It computes contextual SNR using KL divergence from the uniform action distribution; since SNR = log|A| - entropy, high entropy means low SNR.
  • When entropy passes a threshold γ, CAPS runs Metropolis-Hastings search over action chunks: keep a prefix, resample a suffix, then accept or reject the candidate with a power-distribution likelihood ratio.
  • When entropy stays below γ, it uses greedy execution to save inference cost.

Results

  • RoboTwin 1.0 with π0: CAPS reached 47.4% average success, compared with 32.2% for π0 and 41.3% for π0 + TACO; gains were +15.2 points over π0 and +6.1 points over TACO.
  • RoboTwin “Dual Bottles Pick Hard”: success rose from 48.0% with π0 to 61.0% with CAPS; TACO reached 52.0%.
  • Simpler-WindowX: CAPS reached 60.5% average success, compared with 48.0% for π0, 55.5% for π0 + TACO, 42.7% for SpatialVLA, 31.3% for RoboVLM, 16.0% for Octo, and 1.1% for RT-1-X.
  • Simpler-WindowX “Carrot on Plate”: CAPS scored 61.0%, compared with 42.0% for π0 and 52.0% for π0 + TACO.
  • The theory section claims that with α = 2 and single-step drift ε = 0.1, ideal power-distribution sampling gives a 10× effective-horizon extension; with finite MCMC steps, the bound includes a residual O(ρ^N) sampling-bias term.
  • The excerpt says CAPS was also tested on Libero-long, but it does not provide aggregate Libero-long numbers in the visible text.

Link