Recoleta
GitHub
Source note

Evo-Depth: A Lightweight Depth-Enhanced Vision-Language-Action Model

Authors
Tao Lin, Yuxin Du, Jiting Liu, Nuobei Zhu, Yunhe Li, Yuqian Fu, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot ManipulationImplicit DepthMulti View RgbGeneralist Robot Policy
Open arXivMarkdown source

Summary

Evo-Depth is a 0.9B-parameter VLA model that adds RGB-derived depth cues to improve robot manipulation that needs accurate spatial control. It claims higher simulation and real-world success than larger VLA baselines, with lower memory use and faster inference.

Problem

  • 2D-only VLA policies lose accuracy on grasping, placement, and object-object interaction tasks that need depth and relative position.
  • Depth cameras, point clouds, and heavy geometry models add hardware, latency, memory cost, and noise sensitivity.
  • The paper targets deployable manipulation policies that get spatial cues from existing multi-view RGB cameras.

Approach

  • IDEM reads multi-view RGB images and extracts compact latent depth features using a 0.13B plain Vision Transformer initialized from an any-view depth encoder.
  • Early IDEM layers attend within each camera view; later layers mix within-view and cross-view attention so the model can relate objects across views.
  • A pretrained InternVL3-1B vision-language backbone encodes images and the instruction; the paper keeps the first 14 language layers for control.
  • SEM turns IDEM depth features into FiLM-style scale and shift terms, then modulates the vision-language tokens instead of concatenating large 3D features.
  • A flow-matching Diffusion Transformer action expert predicts future actions, trained with three stages: SEM/action expert, then IDEM alignment, then end-to-end tuning.

Results

  • Meta-World: Evo-Depth reaches 84.4% average success with 0.9B parameters; Evo-1 reports 80.6% at 0.8B, and RoboTron Mani reports 77.7% at 4B.
  • VLA-Arena: Evo-Depth reaches 41.1% total success; OpenVLA-OFT reports 39.9% at 7B, OpenVLA reports 39.6% at 7B, and UniVLA reports 38.7% at 7B.
  • LIBERO: Evo-Depth reaches 95.4% average success with no robotic-data pretraining; UniVLA reports 95.2% at 7B, and DepthVLA reports 94.9% at 4.1B.
  • LIBERO-Plus: Evo-Depth reaches 69.6% average success; the next listed baseline, π0-Fast, reports 53.6%.
  • Real-world tests: Evo-Depth reports 90% average success across three tasks, 3.2 GB GPU memory use, and 12.3 Hz inference, with the smallest model size among compared methods.

Link