Fast Visuomotor Policy for Robotic Manipulation

TL;DR

This work introduces Energy Policy, a fast visuomotor policy for robotic manipulation that natively predicts multimodal continuous actions in a single forward pass. By optimizing an energy score with an Energy MLP inside a decoder-only Transformer, the approach learns multimodal action distributions without iterative sampling, enabling parallel decoding and fast inference. Empirical results across Robomimic, Franka Kitchen, MimicGen, and real-world tasks show Energy Policy matches or surpasses state-of-the-art baselines (including diffusion methods and CARP) while achieving substantial speedups (up to tens of times faster in some tasks). The method demonstrates strong multi-task and real-time capabilities, highlighting its potential for resource-constrained, real-time robotic manipulation.

Abstract

We present a fast and effective policy framework for robotic manipulation, named Energy Policy, designed for high-frequency robotic tasks and resource-constrained systems. Unlike existing robotic policies, Energy Policy natively predicts multimodal actions in a single forward pass, enabling high-precision manipulation at high speed. The framework is built upon two core components. First, we adopt the energy score as the learning objective to facilitate multimodal action modeling. Second, we introduce an energy MLP to implement the proposed objective while keeping the architecture simple and efficient. We conduct comprehensive experiments in both simulated environments and real-world robotic tasks to evaluate the effectiveness of Energy Policy. The results show that Energy Policy matches or surpasses the performance of state-of-the-art manipulation methods while significantly reducing computational overhead. Notably, on the MimicGen benchmark, Energy Policy achieves superior performance with at a faster inference compared to existing approaches.

Figures (12)

• Figure 1: Left: Our method achieves strong results in three settings: single-task (top), multi-task (middle), and real-world (bottom). Each row displays rollouts from representative tasks within each scenario. Right: The bubble chart illustrates a representative comparison on PushT chi2023diffusion, showing that our method achieves state-of-the-art success rates while being faster in inference than the strongest baselines. Bubble sizes indicate model parameter counts, demonstrating that our approach delivers competitive performance with significantly smaller models.
• Figure 2: Comparisons of existing Polies. (a) Autoregressive policy predicts discrete tokens in an autoregressive manner. (b) L1/L2 policy predicts continuous actions but struggles with multimodal distribution modeling. (c) Diffusion policy generates multimodal continuous actions through multiple denoising steps. (d) Our Energy Policy produces multimodal continuous actions in a single forward pass.
• Figure 3: Overview of Energy Policy. (a) The architecture of the energy policy primarily consists of a transformer decoder and an energy MLP. The transformer decoder takes observations and learnable action tokens as input, producing a sequence of vectors $\{z_t\}_{t=1}^H$. The energy MLP then predicts the action sequence $\{\hat{a}_t\}_{t=1}^H$ by taking $\{z_t\}_{t=1}^H$ as input, conditioned on noise samples. (b) The energy loss is computed based on two sampled actions, $\hat{a}_t^1$ and $\hat{a}_t^2$, along with the ground-truth action $a_t$. These two action samples are generated from the same $z_t$ using different noise inputs. (c) The energy MLP is composed of several residual blocks, each incorporating adaLN for noise injection and modulation.
• Figure 4: Real-World Experiment Setup. (a) Left: equipment and objects used in the static tasks, with each item labeled by a circled number in the upper-right corner. Middle: desired manipulation behavior for Task Rabbit. Right: desired manipulation behavior for Task Cup . (b) Left: equipment and objects used in the dynamic task, with each item labeled by a circled number in the upper-right corner. Right: desired manipulation behavior for Task Catch.
• Figure 5: Rollout of the first 40 steps of the samples.