VLA Gallery / All models / Flow·Diffusion VLA
Evo-1
MINT-SJTU (GitHub organization name) · 2025-11†
VLAFlow·Diffusion VLA0.77Bflow matching
Reports Meta-World 80.6% and RoboTwin 37.8% at 0.77B without robot-data pretraining, higher than SmolVLA·π0.
Paper
Evo-1: Lightweight Vision-Language-Action Model with Preserved Semantic AlignmentarXiv 2511.04555
Overview
- Category
- Flow·Diffusion VLA
- Type
- VLA
- Released
- 2025-11†
- Organization
- MINT-SJTU (GitHub organization name)
- Open weights
- Yes
- Parameters
- 0.77B
Architecture
- Backbone LLM
- Qwen2.5-0.5B · InternVL3-1B config
- Vision encoder
- InternViT-300M
- Action head
- flow matching DiT (flow matching)
- Chunk
- 50 steps
- Control rate
- Not disclosed
- Architecture (original summary)
- InternVL3-1B (InternViT-300M + Qwen2.5-0.5B) + flow matching Diffusion Transformer action head; chunk H=50
Training
- Pretraining accelerators
- N/A (no pretraining on robot data)
- Pretraining duration
- N/A
- Pretraining steps & batch
- N/A
- GPU/TPU-hours
- Not disclosed
- Fine-tuning
- A100×8; Stage 1 10k steps · batch 16 (VLM frozen), Stage 2 65k steps · batch 16 (full), lr 1×10⁻⁵; total time not disclosed
- Data type
- Simulation + real-robot teleoperation (per task)
- Data scale
- Meta-World 50 demos per task×50 tasks, RoboTwin 50 trajectories per task, real robot 100 teleoperation demos per task
- Main datasets
- Meta-World, LIBERO, RoboTwin, in-house xArm6 data
- Tasks & evaluation
- Meta-World 80.6% (SmolVLA 68.2%), LIBERO 94.8%, RoboTwin 37.8% (π0 30.9%), xArm6 real-robot 4 tasks 78%
- Robot
- xArm6 (single arm) real robot; RoboTwin bimanual simulation
Inference
- Inference device & speed
- RTX 4090d 2.3 GB, 16.4 Hz; Jetson Orin deployment guide (2026-04, README)
- Inference memory
- 2.3 GB (RTX 4090d)
- Weights bf16 (calc.)
- 1.5 GB
- Where it runs
- Not disclosed
Reinforcement learning & base
- Reinforcement learning
- Not used · Imitation learning (behavior cloning) only
- Base control
- Fixed · No base control (xArm6 arm·RoboTwin bimanual sim)
Notes
- Key features
- Reports Meta-World 80.6% and RoboTwin 37.8% at 0.77B without robot-data pretraining, higher than SmolVLA·π0. 16.4 Hz with 2.3 GB on RTX 4090d.
- Notes
- Verification notes only attempted the announcement month, which was unverifiable; other figures unverified. Announcement month unverifiable from arXiv date line† (first README entry 2025-11-06). Included as EVO1 in the LeRobot native policy list. Memory-saving evo1-flash branch.
- Verification
- Not verified