ICRA 2026

Beyond the Teacher: Leveraging Mixed-Skill
Demonstrations for Robust Imitation Learning

S. Saharsh* Shubham Sonkar* Pushpak Jagtap Ravi Prakash

* Equal contribution

Cyber Physical Systems, Indian Institute of Science (IISc), Bengaluru, India
FoCAS Lab HiRO Lab CPS IISc
Paper (Soon) Video Code
Novel IL pipeline for imitation learning with expert-like behavior

Figure 1. Novel IL pipeline for imitation learning with expert-like behavior using imperfect demonstrations. Results from standard IL methods are shown in red and from our proposed pipeline are shown in green.

Abstract

Achieving expert-like robotic task execution in dynamic environments typically requires extensive, high-quality expert demonstrations, a significant bottleneck for real-world deployment. We present a novel learning framework that overcomes this data dependency, enabling robots to perform complex periodic tasks with expert-like proficiency, even when learning from naive demonstrations.

Our two-stage pipeline first selects a representative demonstration based on user-defined information-aware task intention scores. This single best demo is then used to extract a canonical motion shape via Periodic Dynamic Movement Primitives (DMPs). Finally, a Long Short-Term Memory (LSTM) network refines the entire set of demonstrations, leveraging a multi-objective score that combines the canonical shape with mutual information and other task quality metrics.

The proposed approach is demonstrated on a Franka Research 3 robot performing phasic tasks across three contrasting domains: wiping in human assistive services, weaving in the textile industry, and pick-and-place operations for warehouse automation. Code available at: https://github.com/FocasLab/ICRA-IL-2026.

Method

Control flow diagram of the proposed approach

Figure 2. Control flow diagram of the proposed approach showing the two-stage pipeline.

Demo Utility Score

We employ a multi-objective scoring mechanism derived from task-intentional prompts (Do's and Don'ts of the task). The Demo Utility Score U(xi) quantifies the quality of each raw demonstration based on mutual information between states and actions, as well as kinematic and geometric characteristics including symmetry, closure, and smoothness losses.

Canonical Motion Extraction with Periodic DMP

The expert-like trajectory is selected using the demo utility score, and Periodic Dynamic Movement Primitives (DMPs) generate a canonical reference for the desired motion. Rollouts from the learned Periodic DMP estimate a state visitation value map over the task space using kernel density estimation (KDE).

Refinement with LSTM

A Long Short-Term Memory (LSTM) network refines the entire set of demonstrations by learning trajectory corrections. The LSTM leverages the multi-objective score combining the canonical shape with mutual information and task quality metrics to produce expert-like trajectories from imperfect demonstrations.

Results

Our approach achieves expert-like performance on complex periodic tasks, even when trained on as few as four demonstrations.

Noisy vs Cleaned Demonstrations

Comparison of noisy raw demonstrations versus cleaned demonstrations produced by our two-stage pipeline across three periodic task domains. Trajectory projection visualized in the XY plane.

Raw Noisy Demonstrations

Noisy wiping trajectories

Wiping Task (WP)

Noisy pick-and-place trajectories

Pick-and-Place Task (PnP)

Noisy weaving trajectories

Weaving Task (WV)

Cleaned Demonstrations (Ours)

Cleaned wiping trajectories

Wiping Task (WP)

Cleaned pick-and-place trajectories

Pick-and-Place Task (PnP)

Cleaned weaving trajectories

Weaving Task (WV)

IL Model Rollouts

Comparison of rollout trajectories from four imitation learning models trained on noisy and cleaned demonstration data.

Trained on Noisy Data

BC

Traj-BC

ILEED

Neural-ODE

Trained on Cleaned Data

BC

Traj-BC

ILEED

Neural-ODE

Video Results

Demo Video Thumbnail

Full demonstration video showing hardware results from standard IL methods (red) vs. our proposed pipeline (green).

BibTeX

@inproceedings{saharshsonkar2026beyond,
  author    = {S. Saharsh and Shubham Sonkar and Pushpak Jagtap and Ravi Prakash},
  title     = {Beyond the Teacher: Leveraging Mixed-Skill Demonstrations for Robust Imitation Learning},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026},
}