radeonvla_reflex_smolvla_1k_50k

RadeonVLA-Reflex SmolVLA Physical-1K — cumulative 50K

This is the cumulative 50,000-step intermediate checkpoint of RadeonVLA-Reflex. It was trained on the 1,000-episode Physical-1K dataset, not on Physical-2K. It is published as an auditable intermediate model and does not replace the project's existing 20K release.

Provenance

• Base model: lerobot/smolvla_base
• Frozen base revision: c83c3163b8ca9b7e67c509fffd9121e66cb96205
• Training data: a3124371940/radeonvla_reflex_physical_1k
• Cumulative training steps: 50,000
• Checkpoint path at evaluation: ...continue_20k_to80k_v2/checkpoints/030000/pretrained_model
• Pre-publication checkpoint tree SHA256: 466e3c57aacca4e1a479929bbcccc4dcaad319b2fb3295b74e303d736e1a9a3f
• model.safetensors SHA256: 712fcb78449cfc84895d496b287df17ae6b694ecbffc7ab45395fdacbd65d029

The checkpoint includes its policy pre/post-processors and the VLM tokenizer and processor assets required for offline architecture reconstruction.

Inputs and outputs

• Images: world and wrist RGB cameras, 320×240
• State: 9-D Franka qpos (panda_joint1..7, panda_finger_joint1..2)
• Language: a collected fruit-and-bowl instruction
• Action: 9-D absolute joint-position target in the same joint order
• Control frequency: 20 Hz

Intermediate evaluation

The bound evidence file is evaluation/evaluation.physical_1k_cumulative_50k_collected_five_fruit.json. It uses the exact collected language, seeds 50020–50024, one rollout per fruit, no retries, and a 600-step cap.

This five-rollout smoke test is a small intermediate measurement. It is not the formal 20-task score and should not be interpreted as a 60% success claim over the full task suite.

Limitations

The model is simulation-only and limited to the registered five-fruit/four-bowl Genesis scene. Real-robot transfer has not been evaluated. The upstream base repository did not declare license metadata at the frozen revision, so this card uses Hugging Face's other marker rather than inventing a permissive license.