InternVLA-A1.5 发布四个 ModelScope 检查点

InternVLA-A1.5 brings latent foresight to robot control, with four checkpoints now available on ModelScope. 🚀 🤖 https://t.co/x2j0OVrFfz 📄 https://t.co/kqNChLMYBw 🏆 Reported best overall results across six simulation benchmarks, including 98.9 on LIBERO, 84.8 on LIBERO-Plus, and 93.2 on RoboTwin 2.0. 📦 Model series: a 2.69B base checkpoint, plus task-specific versions for LIBERO, RoboTwin 2.0, and DOMINO. 🔮 Latent foresight: learnable tokens capture task-relevant future dynamics from a frozen video model, helping the robot anticipate what comes next without generating pixels. ⚡ Practical control: the training-only video branch is removed at inference, while a lightweight action expert generates continuous actions through flow matching. Trained on 1.2M robot episodes, 861M frames, and 3M multimodal samples. License: CC BY-NC-SA 4.0. https://video.twimg.com/ext_tw_video/2078021334320173056/pu/vid/avc1/1280x720/gnlFiVpY1Ya4bUkK.mp4?tag=12