PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking
PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking
This paper presents PACE, a training-free method for selecting execution horizons in action-chunking robot policies by identifying low-speed valleys in predicted speed profiles. Evaluated on the 50-task RoboTwin2.0 benchmark, PACE improves average success from 57.8% to 64.2% over fixed-horizon baselines and outperforms adaptive execution strategies AAC and AutoHorizon. On real robots, it raises average task scores from 60.7 to 77.7 and success rates from 50.7% to 70.4%. The method is demonstrated on GR00T N1.7 across the LIBERO benchmark and code is available at https://niejunnan25.github.io/pace-project/.
PACE introduces a training-free method that selects online execution horizons using low-speed valleys in predicted action profiles, achieving an average success rate increase from 57.8% to 64.2% on the 50-task RoboTwin2.0 benchmark and improving real-robot task scores from 60.7 to 77.7.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
task score
60.7–77.7
View original evidence
tegies, supporting its applicability across base policies, benchmarks, and action representations. On real robots, PACE improves the average task score from 60.7 to 77.7 and success rate from 50.7% to 70.4% over full-chunk execution. Code is available at https://niejunnan25.github.io/pace-project/.
Open source S5Reported success rate
50.7%
Reported trials
View original evidence
tegies, supporting its applicability across base policies, benchmarks, and action representations. On real robots, PACE improves the average task score from 60.7 to 77.7 and success rate from 50.7% to 70.4% over full-chunk execution. Code is available at https://niejunnan25.github.io/pace-project/.
Open source S5Reported success rate
70.4%
Reported trials
View original evidence
tegies, supporting its applicability across base policies, benchmarks, and action representations. On real robots, PACE improves the average task score from 60.7 to 77.7 and success rate from 50.7% to 70.4% over full-chunk execution. Code is available at https://niejunnan25.github.io/pace-project/.
Open source S5
- code: Available Artifact link Open source S5
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
rates the valley threshold from each task's existing training demonstrations, without task-specific fixed-horizon tuning. On the 50-task RoboTwin2.0 benchmark, PACE improves average success from 57.8% to 64.2% over the strongest fixed-horizon baseline and outperforms the adaptive execution baselines AAC and AutoHorizon. With GR00T N1.7 on LIBERO, PACE achieves the highest overall success rate among the evaluated stra
Open source S4
tegies, supporting its applicability across base policies, benchmarks, and action representations. On real robots, PACE improves the average task score from 60.7 to 77.7 and success rate from 50.7% to 70.4% over full-chunk execution. Code is available at https://niejunnan25.github.io/pace-project/.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org