Official implementation of DexLatent for the paper XL-VLA: Cross-Hand Latent Representation for Vision-Language-Action Models (CVPR 2026 Highlight ✨).

News!

• 2026.04.08: Selected as a CVPR 2026 Highlight ✨.
• 2026.02.27: Released code and project website.
• 2026.02.22: Accepted to CVPR 2026.

Inference and Visualization

Run with our pretrained checkpoint:

uv run -m HandLatent.infer

By default, inference reads Dataset/demo.npz and visualizes:

• source trajectory (origin)
• four decoded trajectories (xhand, ability, inspire, paxini)

Train

Run the training script with:

uv run -m HandLatent.train

Checkpoints are written to:

• Checkpoints/<timestamp>/checkpoint_epoch_XXXX.pt