| CARVIEW |
EfficientFlow: Efficient Equivariant Flow Policy Learning for Embodied AI
AAAI 2026
Abstract
Generative modeling has recently shown remarkable promise for visuomotor policy learning, enabling flexible and expressive control across diverse embodied AI tasks. However, existing generative policies often struggle with data inefficiency, requiring large-scale demonstrations, and sampling inefficiency, incurring slow action generation during inference.
We introduce EfficientFlow, a unified framework for efficient embodied AI with flow-based policy learning. To enhance data efficiency, we bring equivariance into flow matching. We theoretically prove that when using an isotropic Gaussian prior and an equivariant velocity prediction network, the resulting action distribution remains equivariant, leading to improved generalization and substantially reduced data demands.
To accelerate sampling, we propose a novel acceleration regularization strategy. As direct computation of acceleration is intractable for marginal flow trajectories, we derive a novel surrogate loss that enables stable and scalable training using only conditional trajectories.
Across a wide range of robotic manipulation benchmarks, the proposed algorithm achieves competitive or superior performance under limited data while offering dramatically faster inference. These results highlight EfficientFlow as a powerful and efficient paradigm for high-performance embodied AI.
How It Works
To achieve both high efficiency and strong generalization, EfficientFlow introduces two key innovations: Equivariant Flow Matching to handle geometric symmetries, and Flow Acceleration Upper Bound (FABO) to straighten the flow trajectory for faster sampling.
The Geometric Intuition. To handle rotational symmetries, we enforce Equivariance in the flow field (arrows), ensuring that rotating the observation (o → go) results in a correspondingly rotated action trajectory (blue path). Simultaneously, Acceleration Regularization straightens the flow, enabling fast, single-step generation.
EfficientFlow Architecture. At each decision step, the policy takes the past two observations o and a noisy action sequence x0 as input. This information is processed by the equivariant Flow Matching network to generate five candidate action trajectories. The trajectory that exhibits the minimum Euclidean distance to the previously predicted trajectory is then selected for execution, ensuring a smooth and coherent action sequence.
Experiment
EfficientFlow generalizes across 12 challenging tasks in MimicGen.
Quantitative Results
Comparison against SOTA. We report the success rates of 12 MimicGen tasks using 100, 200, and 1000 demos, respectively.
BibTeX
@inproceedings{chang2026EfficientFlow,
author={Chang, Jianlei and Mei, Ruofeng and Ke, Wei and Xu, Xiangyu},
title={EfficientFlow: Efficient Equivariant Flow Policy Learning for Embodied AI},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},
year={2026}
}