RQ1
Can one policy offer control over execution quality?
A task can be completed in many ways, with different levels of speed, smoothness, and safety. How can we learn from this diversity to let users choose the quality they want?
TL;DR Success is only the beginning. QCP controls how a robot performs a task, including its speed, stability, and motion quality, through quality conditioning and iterative hindsight relabeling.
Introduction
Roll out→Measure→Relabel→Train↻k = 0
Select a quality
Smoothness
Turn to adjust
Smooth
Dial up.
Collect more
high-quality rollouts.
Motivation
Successful demonstrations contain many ways to move.
We want to choose the quality, and improve beyond what was demonstrated.
RQ2
Demonstrations cover a limited range.
Selecting a behavior within the initial data is only the first step. Can a controllable policy also learn to reach higher quality beyond those demonstrations?
Method
1. Condition on quality.
Measure each successful trajectory and quantize its score into equal-frequency bins. Train the policy on all successful trajectories, conditioned on their quality labels. At inference time, request a label to select the execution behavior.
Loading method diagram…
2. Expand the controllable range.
Collect rollouts with the highest label and keep successes. Measure their achieved quality, recompute labels for the entire accumulated dataset, and retrain. Repeat as better experience enters the data.
Loading method diagram…
Result
Better quality. Higher success.
QCP improves execution quality and task success together over iterations. It learns from all successful rollouts, using quality labels to select better behavior.
Success rate
QCPSORLSCIQL+BC-QBC
Earlier Later
Robomimic Square
The graph reports aggregate results across three seeds. Videos show representative executions, not individual iterations.
Beyond the initial best.
New rollouts shift the quality distribution beyond the best initial demonstrations. Repeated collection and relabeling expand the range of behaviors the policy can achieve.
Improve multiple qualities together.
One policy is conditioned on six quality attributes at once. Across four tasks, QCP improves most attributes over both initial BC and BC trained on accumulated rollouts.
Lift Pot
Stack
Handover
Valve
QCPBC (accumulate)BC (stem)
Quality tracks the requested level.
q₀q₁q₂q₃ (best)
Changing the quality token selects different execution behaviors from one policy. This ordering is preserved as iterative training improves the achievable quality.
Completion time
Action smoothness
EEF smoothness
EEF speed
Translation path
Rotation path
Video Wall
BibTeX
@misc{kang2026qcp,
title = {Learning Quality-Controllable Robot Policies
with Iterative Hindsight Relabeling},
author = {Kang, Sinjae and Shin, Dong Kyu and Kim, Jaehong and Lee, Kamin
and Kim, Hyoeun and Wang, Kaixin and Zhao, Li
and Ye, Seonghyeon and Lee, Kimin},
year = {2026},
eprint = {2610.37393},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37393}
}