X-Imitator: Spatial-Aware Imitation Learning via Bidirectional Action-Pose Interaction

Summary

Effectively handling the interplay between spatial perception and action generation remains a critical bottleneck in robotic manipulation.

Abstract Summary

Effectively handling the interplay between spatial perception and action generation remains a critical bottleneck in robotic manipulation. Existing methods typically treat spatial perception and action execution as decoupled or strictly unidirectional processes, fundamentally restricting a robot's ability to master complex manipulation tasks. To address this, we propose X-Imitator, a versatile dual-path framework that models spatial perception and action execution as a tightly coupled bidirectional loop. By reciprocally conditioning current pose predictions on past actions and vice versa, this framework enables continuous mutual refinement between spatial reasoning and action generation. This joint modeling exactly mimics human internal forward models. Designed as a modular architecture, the system can be seamlessly integrated into various visuomotor policies. Extensive experiments across 24 simulated and 3 real-world tasks demonstrate that our framework significantly outperforms both vanilla policies and prior methods utilizing explicit pose guidance. The code will be open sourced.

Key Points

Introduces X-Imitator, advancing imitation-learning research.
Detailed evaluation demonstrates effectiveness on benchmarks and real-world scenarios.
Provides insights into methodology and architectural choices.

Method

In this section, we formally present X-Imitator, a spatial-aware imitation learning framework with a bidirectional action-pose interaction mechanism, designed to explicitly incorporate spatial information into modern visuomotor policies. Let $\mathbf{O}_{t}$ denote the visual observation at time step $t$, which typically includes RGB-D images and robot proprioception.

Simulation Experiments

In this section, we empirically evaluate X-Imitator across diverse tasks in several simulation benchmarks. We aim to answer the following research questions: Can X-Imitator improve the performance of various visuomotor policies by leveraging bidirectional action-pose interaction? Does the bidirectional action-pose interaction mechanism lead to better performance than other interactive or non-interactive alternatives?

How does the conditional trajectory length in bidirectional action-pose interaction influence the policy performance? To comprehe...

Simulation Experiments — X-Imitator: Spatial-Aware Imitation Learning via Bidirectional Action-Pose Interaction — Figure 2 : Feature fusion in action branch. Left: Add fusion in X-DP3. Right: Concat fusion in X-ACT. The visual feature $F_{t}^{vis}$ is extracted by the perception module of each base method. The pose trajectory feature $F_{t}^{pose}$ is obtained by: MLP on flattened $P_{t-1}$ for X-DP3 and X-RISE; self-atten on $P_{t-1}$ + cross-atten with $F_{t}^{vis}$ for X-ACT.

Real-World Experiments

In this section, we aim to answer the following research question. Does the advantage of X-Imitator demonstrated in simulation environments translate to real-world scenarios? Our robot platform consists of a Flexiv Rizon 4 robotic arm with a Dahuan AG-95 gripper. A global Intel RealSense D435 RGB-D camera is positioned in front of the robot, perceiving the robot workspace during manipulation.

Following [ 18 ] , we use sigma 7 haptic device to teleoperate the robot to gather 50 expert demonstrations for each task. As illustrated in Fig.

Real-World Experiments — X-Imitator: Spatial-Aware Imitation Learning via Bidirectional Action-Pose Interaction — Figure 3 : Overview of simulated tasks. The first 3 and 4 tasks are from Adroit and Dexart, respectively. The middle 10 tasks with a red robot arm are from MetaWorld, and the last 7 tasks with dual arms are from RoboTwin 2.0.

Sources and Demonstrations

Figures are reproduced from Xiong et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.