Liquid Networks with Mixture Density Heads for Efficient Imitation Learning

2026-03-27

AI Summary

This work addresses key challenges in imitation learning—namely, policy redundancy, slow inference, and inadequate modeling of multimodal action distributions—by introducing, for the first time, a compact and efficient recurrent policy that integrates liquid neural networks with a mixture density output head. Evaluated under shared backbones and identical training conditions, the proposed approach significantly enhances data efficiency and robustness, particularly excelling in low- to medium-data regimes. Compared to diffusion-based policies, it reduces model parameters by approximately 50%, decreases offline prediction error by a factor of 2.4, and accelerates inference speed by 1.8×, demonstrating superior sample efficiency and strong deployment potential across a range of tasks.

Abstract

We compare liquid neural networks with mixture density heads against diffusion policies on Push-T, RoboMimic Can, and PointMaze under a shared-backbone comparison protocol that isolates policy-head effects under matched inputs, training budgets, and evaluation settings. Across tasks, liquid policies use roughly half the parameters (4.3M vs. 8.6M), achieve 2.4x lower offline prediction error, and run 1.8 faster at inference. In sample-efficiency experiments spanning 1% to 46.42% of training data, liquid models remain consistently more robust, with especially large gains in low-data and medium-data regimes. Closed-loop results on Push-T and PointMaze are directionally consistent with offline rankings but noisier, indicating that strong offline density modeling helps deployment while not fully determining closed-loop success. Overall, liquid recurrent multimodal policies provide a compact and practical alternative to iterative denoising for imitation learning.

Problem

• imitation learning
• efficient policy
• sample efficiency
• low-data regime
• policy modeling

Innovation

• Liquid Neural Networks
• Mixture Density Heads
• Imitation Learning
• Sample Efficiency
• Diffusion Policies