InternRobotics/InternVLA-A-series

What it solves

InternVLA-A1.5 addresses the challenge of compositional generalization in robotic manipulation by unifying three critical capabilities—vision-language understanding, visual foresight (predicting future dynamics), and action generation—into a single policy.

How it works

The system uses a Qwen3.5-2B VLM as its backbone, adding a lightweight unified action expert via shared full-attention layers. During training, it employs learnable foresight tokens to query future dynamics from a multimodal context, supervised by a frozen WAN2.2-5B video generation model. For actual deployment, the video branch is discarded to maintain low latency, and the model predicts continuous action chunks using flow matching.

Who it’s for

It is designed for robotics researchers and developers working on robotic manipulation, specifically those looking to improve how robots generalize tasks across different simulation benchmarks (like RoboTwin, LIBERO, and SimplerEnv) and real-world settings.

Highlights

• Unified Architecture: Combines understanding, foresight, and action in one model.
• Latent Foresight: Uses a video generation model during training to teach the policy about future task dynamics.
• Efficient Inference: Discards the heavy video branch at runtime to ensure practical deployment latency.
• Broad Compatibility: Supports fine-tuning on LeRobot V2.1 datasets and various simulation environments.