Improving Robotic Imitation Learning via Trajectory Standardization

Yang, Licheng; Qian, Lingfeng; Zheng, Fei; He, Yonghao; Sui, Wei; Li, Shuangshuang; Su, Hu

Abstract

Imitation learning for robotic manipulation relies on large sets of human demonstration trajectories, which are often noisy and temporally irregular due to variable operator speed, intermittent pauses, and inconsistent action density. A common preprocessing strategy is time-uniform downsampling to shorten sequences, but it cannot effectively remove speed-induced non-uniformity or redundant pauses. This mismatch degrades data quality and hinders policy learning. To address this issue, we propose Information-Standardized Trajectory Resampling (ISR), an offline preprocessing method for effective imitation learning. ISR resamples each trajectory by enforcing approximately equal information distance between adjacent points. Specifically, we map trajectories onto an information-modulated Riemannian manifold and perform geodesic-equidistant parameterization. We construct an information-intensity field from velocity and acceleration norms: the velocity term removes small-motion redundancy, while the acceleration term preserves high-curvature and fine-manipulation phases. We evaluate ISR on three real-world manipulation tasks with mainstream imitation learning policies. Compared with the baseline time-uniform 3x downsampling, ISR improves task success rates by about 25%, remains robust across datasets collected from different operators, and reduces both dataset size and training cost.

Method

Information-Standardized Trajectory Resampling (ISR) is an offline preprocessing method that resamples raw teleoperated demonstration trajectories so that consecutive retained points carry approximately equal kinematic and dynamic information. ISR comprises three steps, as illustrated in the figure below.

Step 1 — Information Intensity Field. From the end-effector position sequence, ISR computes per-frame velocity and acceleration via finite differences, then combines them into a scalar information-intensity field weighted by two complementary terms. The velocity term captures the instantaneous displacement rate and identifies low-value segments such as unintentional pauses and slow transits. The acceleration term encodes curvature-induced centripetal acceleration during arc-like movements and tangential deceleration before contact-rich manipulation phases, serving as a kinematic proxy for force-sensitive operator intent.

Step 2 — Riemannian Manifold Mapping. Using the information-intensity field as a conformal factor, ISR warps each trajectory onto a one-dimensional Riemannian manifold. Under this metric, high-activity regions are stretched while low-activity regions are compressed, so that equal geodesic arc lengths correspond to equal amounts of kinematic–dynamic information. The geodesic distance between two trajectory points is approximated in discrete form using the secant distance for the velocity term and the accumulated acceleration sum for the acceleration term.

Step 3 — Parametric Standardization. ISR selects the resampled index set by minimizing the total squared deviation of each segment's geodesic distance from a prescribed target distance. This geodesic-equidistant optimization is solved exactly via dynamic programming with prefix-sum acceleration, producing a compact, standardized trajectory that serves as a drop-in replacement for any downstream imitation learning pipeline without architectural changes.

Effectiveness

ISR redistributes samples from low-value motion to task-critical phases. Its velocity term compresses pauses, hesitation, and slow low-displacement transit into larger, more uniform steps, reducing operator-speed artifacts and redundant frames; its acceleration term keeps denser coverage around sharp turns, deceleration-to-contact, and fine manipulation, preventing compact but essential actions such as grasping, alignment, and placement from being skipped. This information-balanced resampling improves downstream success while retaining only about 28-40% of the original action points and remaining more robust when demonstrations come from multiple operators.

Evaluation

We evaluate ISR on three real-world manipulation tasks: Place&Cover, Place&Stack, and Push-T. The evaluation contains three studies. Baseline Comparison compares ISR with time-uniform 3x downsampling under two imitation learning policies, π_0.5 and VO-DP. Ablation Study varies the acceleration weight while keeping the velocity term fixed to analyze how dynamic-information preservation affects downstream policy learning. Cross-Operator Robustness tests whether ISR can absorb operator-specific differences in speed, pauses, and teleoperation style when demonstrations are collected from multiple people.

Across the three studies, ISR consistently improves downstream policy learning by redistributing samples according to task-relevant motion information rather than timestamp spacing. The baseline comparison shows broad gains across tasks and policy backbones; the ablation study shows that the acceleration term is essential for contact-sensitive manipulation; and the cross-operator study shows that information standardization reduces style-induced conflicts when demonstrations are collected from multiple operators.

Baseline Comparison

Method	Place&Cover	Place&Stack	Push-T	AVG. (↑)
3x + π_0.5	45.8 (33.3)	48.6 (33.3)	48.9 (33.3)	47.8
ISR + π_0.5	72.2 (28.9)	72.2 (28.2)	71.1 (40.2)	71.8
3x + VO-DP	56.9 (33.3)	66.7 (33.3)	64.4 (33.3)	62.7
ISR + VO-DP	77.8 (28.9)	91.6 (28.2)	71.1 (40.2)	80.2

Success rate is shown in percent; resampling ratio is shown in parentheses.

Acceleration weight ablation on Place and Cover — Ablation Study (Acceleration Weight in Parentheses)

Acceleration weight ablation on Place and Stack — Ablation Study (Acceleration Weight in Parentheses)

Single-operator Place and Stack comparison — Cross-Operator Robustness

Mixed-operator Place and Stack comparison — Cross-Operator Robustness

Limitations

ISR currently operates on end-effector position trajectories. Extending the information-intensity field to joint-space representations could broaden its applicability to robots and tasks where joint-level motion carries important information beyond the end-effector path. In addition, the acceleration weight λ_acc is manually selected for each task, since different manipulation behaviors can have different sensitivity to acceleration-level features. An adaptive mechanism that infers this weight from trajectory statistics would make ISR more general and easier to deploy across new tasks.

BibTeX

@inproceedings{yang2026isr,
  title={Improving Robotic Imitation Learning via Trajectory Standardization},
  author={Licheng Yang and Lingfeng Qian and Fei Zheng and Yonghao He and Wei Sui and Shuangshuang Li and Hu Su},
  booktitle={arXiv:2606.22907},
  year={2026}
}