marker_100, RT-1 backbone warm-startmarker_100 SO-101 teleoperation dataset (97 episodes,
31,291 frames, single_arm(5)+gripper(1) absolute joint-position action).qwen_vl_interface warm-started from StarVLA/Qwen3VL-PI_v3-Bridge-RT_1
(checkpoints/steps_50000_pytorch_model.pt), then frozen for the whole run.
action_model + project_layers are freshly initialized (the RT-1 checkpoint's
action space is 7-dim WidowX/Google-robot EE-delta, incompatible with SO-101's
6-dim absolute joint action, so only the backbone transfers).marker_100), 49h03m wall-clock on a single RTX 3090 24GB.action_dit_loss): 0.049 at step 115,000 (down from 2.92 at step 10;
stabilized around 0.01-0.1 from ~step 4,000 onward).final_model/pytorch_model.pt only. Eleven intermediate
checkpoints (every 10,000 steps) exist locally but were not uploaded (112GB total).1python deployment/model_server/server_policy_gr00t_zmq.py \
2 --ckpt_path final_model/pytorch_model.pt \
3 --port 5555 --use_bf16Gr00t16ServicePolicyClient (--policy_type gr00tn1.6) out of the box — see
examples/SO101_Marker/eval_files/ in the StarVLA repo.