lingbot-va

Why we included this project

LingBot-VA tackles the part of generalist robot control that trips up most policies: predicting what happens next in a scene while deciding what to do. It handles both in one autoregressive video-action world model, interleaving visual dynamics and action inference in a single sequence, so the robot can weigh the consequences of its behavior before acting. The repo is set up for real use, with pretrained checkpoints, post-training code, and cleaned datasets in LeRobot format, plus ready-to-run evaluation on RoboTwin-2.0 and LIBERO and a server-client setup that keeps the model separate from the simulator. That makes it a good fit for robotics researchers and embodied-AI engineers who want to reproduce the results, fine-tune on their own manipulation data, or compare a world-model approach against action-only policies. The dual-stream mixture-of-transformers design with KV cache is also worth a look if inference efficiency on long-horizon tasks matters to you.