evo1

Overview

Second-highest success rate of the 7 (94.5% on LIBERO-Object, RTX 3060), fits comfortably on Orin Nano. No known gating requirement, but not yet individually verified against its HF model card the way smolvla/pi0/vla-jepa were — check before adding to catalog.

Architecture

Vision backbone

InternViT-300M

Language backbone

Qwen2.5-0.5B

Action head

cross-attention DiT (flow-matching)

Action steps per chunk

8

Solver steps

32

Camera views

3 — obs.pixels.image (real, front), obs.pixels.image2 (real, wrist), zero-filled pad slot

Server-side view count is flexible, but the real client (adapters.py:75-89) always sends 3 image slots with image_mask=[1,1,0] — 2 real camera views + 1 masked pad, not 2. Uses raw sim camera key names (obs.pixels.*), NOT the observation.images.* convention other archs use.

Hardware fit

Benchmarks