Demonstrates that Liquid Neural Networks can outperform Diffusion Policies in imitation learning with half the parameters and nearly 2x faster inference.
March 31, 2026
Original Paper
Liquid Networks with Mixture Density Heads for Efficient Imitation Learning
arXiv · 2603.27058
The Takeaway
Diffusion policies have become the standard for robotics, but this work shows that recurrent liquid models are more sample-efficient and robust in low-data regimes. For practitioners, this offers a significantly more efficient path for real-time embodied AI control without the iterative denoising overhead.
From the 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,