random-sequence/flock-robotics-qwen25vl-vla-4b

FLock Robotics VLA - Qwen2.5-VL-3B + action head (~4.06B)

Vision-language-action policy for the FLock AI Arena robotics task. A frozen Qwen2.5-VL-3B-Instruct backbone (LoRA-adapted, merged) with an MLP action head, behaviour-cloned on scripted robosuite Lift trajectories.

flock_robotics_adapter.py exposes load_policy(model_dir, device, dtype) returning a policy whose act(obs) maps an agentview image + instruction to a 7-DoF action. ~4.06B parameters (under the 7B validator cap).