InternVLA-A1

Shanghai AI Lab · 2026-01

VLANew in 20261.8Bflow matching

Coordinates scene understanding, visual prediction generation, and action execution within a single MoT.

Paper

Backbone LLM: InternVL3 2B variant · Qwen3-VL 3B variant (two variants)
Vision encoder: VLM built-in
Action head: flow matching action expert + visual prediction (MoT) (flow matching)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): MoT with 3 experts: understanding (InternVL3 2B variant·Qwen3-VL 3B variant), generation (Cosmos CI8×8 continuous VAE tokenizer, visual prediction), action (chunk flow matching)

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: 700K steps, batch 512, fixed LR 5×10⁻⁵, bf16
GPU/TPU-hours: Not disclosed
Fine-tuning: GPU and time not disclosed; 60,000 steps, batch 128, LR 5×10⁻⁵→5×10⁻⁶
Data type: Simulation·synthetic (InternData-A1) + real-robot teleoperation (AgiBot World)
Data scale: Over 533M frames (synthetic 325M + AgiBot World Beta 208M); full InternData-A1 630k trajectories · 7,433 hours
Main datasets: InternData-A1, AgiBot World (Beta)
Tasks & evaluation: 12 real-world tasks 75.1% (3B variant, π0 60.6%); picking moving ingredients 93.3% (π0 20%); RoboTwin 2.0 Easy 65.0%·Hard 25.4%
Robot: Genie-1, ARX Lift-2, ARX AC One

Inference device & speed: About 13 Hz on RTX 4090 (torch.compile, both variants)
Inference memory: Not disclosed
Weights bf16 (calc.): 6.4 GB
Where it runs: Not disclosed

Reinforcement learning: Not used · Visual prediction + flow matching supervised learning
RL evidence: https://arxiv.org/html/2601.02456v1
Base control: Fixed · No base-movement tasks (bimanual manipulation; no mention of chassis·lift control)

Key features: Coordinates scene understanding, visual prediction generation, and action execution within a single MoT. Trained on a mix of 325M synthetic frames and 208M real-robot frames.
Notes: Borderline case: arXiv v1 2026-01-05 (v2 2026-02-13), but repo created 2025-09-18 (aggregator, secondary), so code·weights may predate 2026. The synthetic 325M row does not appear in the verification reading. HF updated checkpoint (with human video added) reports RoboTwin 2.0 Easy 89.40%·Hard 89.64% under a different protocol. Training total is in frames, so data hours (h) left blank. Parameters (B) is the large variant.
Verification: Confirmed