VLA GalleryAll modelsGallery

VLA Gallery / All models / New in 2026

InternVLA-A1.5

Shanghai AI Lab · 2026-07†

VLANew in 2026Total amount not disclosedNot disclosed

Supervises latent visual prediction tokens with a WAN2.2-5B video teacher during training only, and drops the video branch at inference.

View architecture diagram in the galleryPaper ↗

Paper

InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional GeneralizationarXiv 2607.04988

Overview

Category
New in 2026
Type
VLA
Released
2026-07†
Organization
Shanghai AI Lab
Open weights
Yes
Parameters
Total amount not disclosed (VLM 2B + action expert 460M; WAN2.2-5B teacher used for training only)

Architecture

Backbone LLM
Qwen3.5 2B (Gated DeltaNet hybrid)
Vision encoder
—
Action head
Unified action expert 460M (Not disclosed)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
Qwen-3.5 2B VLM (3 Gated DeltaNet linear attention layers interleaved with 1 full attention layer) + 460M unified action expert; 50 visual prediction tokens (supervised by a frozen WAN2.2-5B video generator during training; video branch removed at inference)

Training

Pretraining accelerators
Not disclosed
Pretraining duration
Not disclosed
Pretraining steps & batch
Stage 1 300K + Stage 2 600K steps, batch 1024, LR 5×10⁻⁵, bf16
GPU/TPU-hours
Not disclosed
Fine-tuning
Post-training with cosine decay; hardware and time not disclosed
Data type
Robot manipulation data (6 sources, real·synthetic split Not stated) + web vision-language (multimodal QA)
Data scale
Robot 1.2M episodes·861M frames; multimodal QA about 3M; sampling robot:multimodal = 0.15:0.85
Main datasets
6 sources (names not stated in note)
Tasks & evaluation
Best overall on 6 benchmarks: LIBERO·RoboTwin 2.0·EBench·SimplerEnv·LIBERO-Plus (zero-shot)·DOMINO (zero-shot) (self-reported); real-world tube insertion 80.7%·tube transfer 74.4%·MOF long-horizon 76.4%
Robot
Not stated in notes

Inference

Inference device & speed
About 0.1 s per inference step on a single RTX 5090
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Local PC

Reinforcement learning & base

Reinforcement learning
Not used · Supervised learning (within the disclosed scope)
Base control
Fixed · No base control (manipulation benchmarks such as RoboTwin 2.0·LIBERO·SimplerEnv and real manipulation)

Notes

Key features
Supervises latent visual prediction tokens with a WAN2.2-5B video teacher during training only, and drops the video branch at inference. Self-reports the best overall results on all 6 simulation benchmarks.
Notes
Corrections applied: weights released (base·RoboTwin·Libero·DOMINO checkpoints on HF·ModelScope, CC BY-NC-SA 4.0). arXiv submission date 2026-07-06† unverifiable because the abs page returned HTTP 429 (month matches arXiv ID 2607). Sampling order fixed at robot 0.15. The author field 'Physical Intelligence Team' is a Shanghai AI Lab team name, unrelated to the company Physical Intelligence.
Verification
Corrections applied

Other models in this category