AsymEEF: Bimanual Manipulation
with Asymmetric End Effectors

Paper and Code will be added with the public release.

Abstract

Teleoperated demonstrations are costly to collect and rarely reusable across end effectors with different kinematics and action spaces. This burden grows in bimanual manipulation, where each end-effector pair may require new data. We present AsymEEF, a framework for executable cross-end-effector demonstration generation and unified bimanual policy learning.

Given a small source dataset, AsymEEF transfers demonstrations to heterogeneous combinations of two-, three-, and five-fingered end effectors without additional teleoperation. Kinematic teacher anchors provide geometric supervision, while sparse real-world X-Calib anchors correct discrepancies in grasp execution. A shared encoder maps wrist poses and actuator states to a morphology-agnostic latent space, and a morphology-conditioned decoder reconstructs target commands. Generated trajectories undergo replay filtering for task-level executability.

The representation also supports a unified bimanual visuomotor policy trained across multiple pairs. Across four simulated and six real-world tasks, AsymEEF improves replay success over adapted transfer baselines, achieves a mean replay-valid expansion gain of 16.6× in simulation, and is evaluated through downstream policy training on real robots.

Framework overview: homogeneous source demonstrations are transferred to heterogeneous pairs, validated by task replay, and used for unified policy learning.
Figure 1. AsymEEF transfers a small set of demonstrations across asymmetric end-effector pairs and learns a unified bimanual policy.

Real-world Demonstrations

Six contact-rich tasks evaluate asymmetric manipulation with XHAND, the G2 gripper, and a fixed tool bracket.

Method

AsymEEF connects demonstration generation and policy learning through a shared EEF-state interface.

Shared EEF-state encoder, latent bimanual policy, and morphology-conditioned decoder.
Figure 2. A shared encoder maps wrist poses and actuator states into a morphology-agnostic latent. The decoder reconstructs target-specific commands. The same interface supports ACT and π0.5 policies.

Cross-EEF Anchoring

Geometry-guided teacher anchors align functional grasp states. Sparse X-Calib anchors capture discrepancies in real grasp execution.

Kinematic teacher anchoring and real-world X-Calib anchoring.
Figure 3. Geometric supervision and physical calibration.

Transfer and Validation

Decoded trajectories are retained only after task replay, including checks for kinematic feasibility, collisions, contact progression, and sustained success.

Source frame and target alignments for AsymEEF, in-hand-only alignment, and DCD alignment.
Figure 4. Target EEF alignment relative to the source grasp frame.

Interactive Cross-EEF Transfer

Choose an end effector for each arm to explore the same bimanual motion across morphologies.

Preview of the bimanual XHAND motion in the interactive 3D viewer.

Loading 3D scene…

Recorded XHAND motion is shown alongside decoded target trajectories. This viewer illustrates kinematic transfer; task success is evaluated separately in the experiments below.

Experimental Results

Evaluating replay and closed-loop policy success.

Demonstration Replay

Success rate (%) ↑

Policy Evaluation

Success rate (%) ↑

Data Expansion

Gain = accepted / source demonstrations; acceptance = accepted / generated. Progressive expansion uses a separate protocol from the direct replay evaluation above.

Across 18 source–target settings, 4,959 of 6,549 generated candidates pass replay. Each source demonstration yields 11.36–19.88 accepted trajectories.

Representation Learning Ablations

Lower is better. The aggregate averages five normalized metrics: Native MAE, Cross MAE, wrist translation error, wrist rotation error, and Real Cross MAE. Full is not best on every individual metric.