ELAN4D adds 4D future tracking to make VLA robot policies 30% more robust

ELAN4D, proposed by a team led by Zeyuan He and including Philip Torr, addresses a key limitation of current Vision-Language-Action (VLA) models: their reactive nature. Most VLA policies directly regress actions from current observations without modeling future dynamics, making them brittle under perturbations. ELAN4D introduces embodiment-centric 4D supervision by extracting future 3D displacement tracks of robot keypoints (joints and end-effector) solely from forward kinematics derived from proprioceptive states. This requires no external trackers or scene reconstruction, keeping preprocessing costs negligible. A plug-and-play auxiliary branch with a lightweight track decoder injects this predictive signal into the action expert during training, while gradient isolation preserves the pretrained vision-language backbone. At inference, the decoder is discarded, so the policy interface remains unchanged.

Extensive experiments on LIBERO, LIBERO-Plus, RoboTwin2.0, and real-world manipulation tasks demonstrate consistent improvements over strong VLA baselines. ELAN4D achieves the best overall performance and shows particularly large gains under out-of-distribution perturbations such as camera angle changes, background shifts, and layout alterations. The method is model-agnostic and can be applied to various VLA architectures. The authors argue that explicitly modeling temporal dynamics through embodiment-centric supervision is a key step toward building more robust and generalizable manipulation policies, especially for deployment in unstructured environments where visual conditions vary significantly.