ASkDAgger: Active Skill-level Data Aggregation for Interactive Imitation Learning

Authors: Jelle Luijkx, Zlatan Ajanović, Laura Ferranti, Jens Kober

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the effectiveness of ASk DAgger through language-conditioned manipulation tasks in both simulation and real-world environments.
Researcher Affiliation Academia Jelle Luijkx EMAIL Department of Cognitive Robotics Delft University of Technology Zlatan Ajanović EMAIL Department of Computer Science RWTH Aachen University Laura Ferranti EMAIL Department of Cognitive Robotics Delft University of Technology Jens Kober EMAIL Department of Cognitive Robotics Delft University of Technology
Pseudocode Yes Algorithm 1: Active Skill-level DAgger (ASk DAgger) [...] Algorithm 2: S-Aware Gating (SAG) [...] Algorithm 3: Foresight Interactive Experience Replay (FIER) [...] Algorithm 4: Prioritized Interactive Experience Replay (PIER)
Open Source Code Yes Code, data, and videos are available at https://askdagger.github.io.
Open Datasets Yes First, we performed active dataset aggregation on the MNIST dataset (Le Cun et al., 1998) [...] The CLIPort benchmark includes seen and unseen task settings.
Dataset Splits Yes The CLIPort benchmark includes seen and unseen task settings. In the unseen setting, test-time commands involve different objects, shapes, or colors than during training.
Hardware Specification Yes The MNIST experiments (Sec. 5.1) and real-world experiments (subsection 5.3 and subsection 5.4) were performed using an RTX 3080 Mobile graphics card. The CLIPort simulation benchmark experiments (Sec. 5.2) were performed using multiple A40 graphics cards on a high-performance computing cluster.
Software Dependencies No The control scheme is implemented using the EAGERx framework (van der Heijden et al., 2024). The Gradio (Abid et al., 2019) interface allows command input via speech or text. Additionally, existing packages such as Torch Uncertainty (Lafage & Laurent, 2024) facilitate uncertainty quantification in this setting.
Experiment Setup Yes For SAG we used mode = sensitivity, σdes = 0.9, Nmin = 15, prand = 0.2 and for PIER α = 1.5, b = 10, β = 1 and λ = 0.5. Each setting involved training ten CLIPort agents without BC pretraining, collecting 300 interactive demonstrations, and evaluating checkpoints every 100 demonstrations.