Research map / Vision-language and multimodal / Robot action and navigation

PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Multimodal prompt-driven manipulation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PACE, a training-free test-time rule that picks each action-chunk execution horizon online from low-speed valleys in the predicted speed profile, improving pi0.5 success on RoboTwin2.0 and real-robot ALOHA/Franka tasks.

Problem

Action-chunking robot policies require choosing an execution horizon (how many predicted actions to run before re-querying), and fixed horizons yield task-dependent, non-monotonic success.

Approach

Select the executed prefix per query by detecting prominent low-speed valleys in the smoothed speed profile of the predicted action chunk and replanning at the earliest accepted valley, falling back to a maximum horizon otherwise; uses only the predicted chunk, no retraining or policy internals.

Key results

• On 50 RoboTwin2.0 tasks, PACE raises average success from 57.8% (best fixed H=25) to 64.2%.
• On real robots, average Score rises from 60.7 to 77.7 and Succ. from 50.7% to 70.4% across three tasks.
• At matched average horizon, PACE beats the matched fixed-horizon schedule on all six representative tasks (gains 1.6 to 10.6 points).

Surprising findings

• Success vs. fixed execution horizon is strongly non-monotonic and task-dependent; e.g., Click bell drops ~40 points around H≈25 and rebounds near H≈35.
• Smoothing the speed profile increased the selected horizon and success, while the hindsight-best fixed horizon varies widely across tasks.

Limitations

• Does not change the base policy's action distribution; cannot fix missing task-completing behavior (failed pencil-case rollout).
• Relies on low-speed valleys reflecting phase boundaries; tasks without such structure or noisy motion profiles may give weaker signals.
• Evaluated only with the pi0.5 policy family.

Experimental setup

Datasets: RoboTwin2.0 RoboChallenge

Benchmarks: RoboTwin2.0 RoboChallenge In-lab Franka place_object_on_plate

Metrics: success rate task score

Baselines: Fixed horizon H=5 Fixed horizon H=25 Fixed horizon H=50 AutoHorizon

Builds on: Diffusion Policy ACT pi0.5

Closest papers by meaning

• FASTER: Rethinking Real-Time Flow VLAs (NeurIPS 2026 · similarity 0.90)
• Real-Time Execution of Action Chunking Flow Policies (NeurIPS 2025 · similarity 0.90)
• Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling (ICLR 2025 · similarity 0.89)
• Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · similarity 0.88)
• PIVOT-R: Primitive-Driven Waypoint-Aware World Model for Robotic Manipulation (NeurIPS 2024 · similarity 0.87)
• SkiP: When to Skip and When to Refine for Efficient Robot Manipulation (NeurIPS 2026 · similarity 0.87)