Reverse Playback Motion:
A Simple Way for Anyone to Obtain High-Quality Demonstration Data

Hyunjun Kim1*, Jaeseog Won1*, Hong-ryul Jung2*, Donghoon Kim3, Minsuk Ko3, Hyungpil Moon1,2
1Intelligent Robotics, Sungkyunkwan University   2Mechanical Engineering, Sungkyunkwan University
3Production Engineering Research Institute, LG Electronics
Int. J. Precision Engineering and Manufacturing (IJPEM), 2026
*Indicates Equal Contribution
Paper Springer Code (Coming Soon)

Hard sub-tasks are teleoperated backward from the goal state, easy ones forward. Time-reversing and re-simulating the reverse segments yields a complete forward demonstration — the operator never has to perform the difficult alignment motion.

Abstract

High-quality robot learning data is essential for advancing Vision-Language-Action (VLA) models, yet collecting it at scale remains the primary bottleneck in their development. This bottleneck is further amplified for manipulation tasks that require tight pose alignment or impose geometric constraints, where small errors lead to task failure, such as peg-in-hole. To address this, we propose Reverse Playback Motion (RPM), a novel data collection framework for robot manipulation. Rather than attempting to precisely place an object into a target configuration, operators instead start from the goal state and perform the reverse motion, which is inherently easier to execute. The collected reverse motions are then reconstructed into valid forward trajectories through time-reversal of the recorded end-effector poses. We evaluate RPM across five constrained manipulation tasks, including both single-stage and multi-stage scenarios, and demonstrate that it enables up to 3 times faster data collection for non-expert operators, while achieving superior policy performance compared to conventionally collected demonstrations.

Overview

Overview of reverse playback motion for single-stage and multi-stage tasks

Numbered frames are key states of each task. A task is split into sub-tasks, and each sub-task is teleoperated in whichever direction is easier: red segments are recorded forward, blue segments are recorded in reverse, starting from the goal. The “Recorded a→b” labels give the order in which the operator actually teleoperated each segment — note that it is not the order the robot will eventually execute.

The full demonstration is then obtained by playing the recorded segments back along the time-ordered trajectory. Single-stage tasks (top) need one reverse-recorded segment; multi-stage tasks (bottom) alternate several forward and reverse segments. Either way the operator skips every hard alignment motion, and what comes out is an ordinary forward demonstration.

Why reverse?

Teleoperation difficulty is asymmetric. Inserting a peg into a hole demands millimeter-level alignment; extracting the same peg starts from a feasible contact configuration and skips the alignment phase entirely. Better still, the geometric boundaries that obstruct the forward motion act as natural guides along the reverse path.

Naively replaying the recorded actions backwards does not work, because the dynamics are not time-symmetric — gravity resists upward motion but assists downward motion, while friction always opposes travel. Reversing the action stream also degrades execution to open loop, and residual errors accumulate until the trajectory diverges.

RPM therefore treats the time-reversed end-effector pose sequence as a reference trajectory and re-executes it in simulation, re-deriving the actions from the tracking error. Both the action stream and the camera images are recorded anew, so every modality stays in sync with a physically valid rollout; only successful regenerations are kept. Forward and reverse segments are then concatenated and augmented at scale with MimicGen.

Up to 3× faster collection

Successful demonstrations collected per 30-minute session, with and without RPM

An expert (Operator A, 100+ hours) and a non-expert (Operator B, under one hour) each collected data for 30 minutes per task. RPM helps both, but the gain is largest for the non-expert and for longer-horizon tasks — exactly where conventional teleoperation is weakest. On Peg-in-Hole, Operator B goes from 12 to 36 demonstrations (3×); on Stacking-and-Peg Assembly, from 5 to 21.

Better policies, same data quantity

We fine-tune GR00T N1.5 on the augmented datasets. The same number of seed demonstrations and the same augmentation factor are used with and without RPM, so any difference reflects data quality, not quantity. Success rates improve on every one of the five tasks.

Without RPM
With RPM

Cube Stacking

Single-stage  (success rate) Cube Stacking Slot-Insertion Peg-in-Hole
Relative task pose32%60%36%
Relative task pose + RPM60%70%50%
Absolute joint angle78%96%56%
Absolute joint angle + RPM96%98%74%
Multi-stage  (success rate, with Temporal Ensemble) Three-Cube Stacking Stacking-and-Peg Assembly
Absolute joint angle76%6%
Absolute joint angle + RPM86%32%

The advantage grows with task horizon.

Where the gain comes from

Trajectory comparison between conventional teleoperation and RPM

Conventional teleoperation oscillates and collides during the final approach — variability that destabilizes the learned policy. Recording from the goal state suppresses that noise: trajectory standard deviation drops from 3.74 to 2.66 (−28.8%) for stacking and 4.04 to 2.77 (−31.5%) for insertion.

BibTeX

@article{kim2026reverse,
  title={Reverse Playback Motion: A Simple Way for Anyone to Obtain High-Quality Demonstration Data},
  author={Kim, Hyunjun and Won, Jaeseog and Jung, Hong-ryul and Kim, Donghoon and Ko, Minsuk and Moon, Hyungpil},
  journal={International Journal of Precision Engineering and Manufacturing},
  pages={1--11},
  year={2026},
  publisher={Springer}
}