| --- |
| license: apache-2.0 |
| tags: |
| - robotics |
| - rlinf |
| - reinforcement-learning |
| - embodied-ai |
| - lingbotvla |
| - robotwin |
| pipeline_tag: robotics |
| --- |
| |
| # RLinf-lingbotvla-click-bell-grpo |
|
|
| This is a LingbotVLA checkpoint post-trained with RLinf GRPO on the RoboTwin `click_bell` task. |
|
|
| The checkpoint has been merged into the pinned RoboTwin SFT base format, so it can be loaded directly as a LingbotVLA model directory. It does not require a separate `runner.ckpt_path` file for evaluation. |
|
|
| ## Base Model |
|
|
| - Upstream base: [`robbyant/lingbot-vla-4b-posttrain-robotwin`](https://huggingface.co/robbyant/lingbot-vla-4b-posttrain-robotwin) |
| - Base revision: `3e0c7c476bde3daaac00f79f3741a292a299f60a` |
| - RL method: RLinf GRPO |
| - RoboTwin task: `click_bell` |
|
|
| ## Result Summary |
|
|
| In our RoboTwin random evaluation, this RLinf-GRPO checkpoint performs better than the pinned RoboTwin SFT checkpoint on the corresponding `click_bell` task. |
|
|
| ## Download |
|
|
| ```bash |
| hf download RLinf/RLinf-lingbotvla-click-bell-grpo \ |
| --local-dir RLinf-lingbotvla-click-bell-grpo |
| ``` |
|
|
| ## Evaluation in RLinf |
|
|
| Use the downloaded directory as both `actor.model.model_path` and `rollout.model.model_path`. Keep `tokenizer_path` pointing to a local Qwen2.5-VL-3B-Instruct directory. |
|
|
| ```bash |
| python examples/embodiment/eval_embodied_agent.py \ |
| --config-path examples/embodiment/config \ |
| --config-name robotwin_click_bell_grpo_lingbotvla \ |
| runner.only_eval=True \ |
| actor.model.model_path=/path/to/RLinf-lingbotvla-click-bell-grpo \ |
| actor.model.tokenizer_path=/path/to/Qwen2.5-VL-3B-Instruct \ |
| actor.model.lingbotvla.config_path=/path/to/RLinf-lingbotvla-click-bell-grpo \ |
| rollout.model.model_path=/path/to/RLinf-lingbotvla-click-bell-grpo \ |
| rollout.model.tokenizer_path=/path/to/Qwen2.5-VL-3B-Instruct \ |
| rollout.model.lingbotvla.config_path=/path/to/RLinf-lingbotvla-click-bell-grpo |
| ``` |
|
|
| For large RoboTwin evaluations, set `env.eval.total_num_envs`, `env.eval.max_episode_steps`, and `env.eval.max_steps_per_rollout_epoch` according to the target protocol. |
|
|
| ## Notes |
|
|
| This repository stores the merged LingbotVLA model weights only. RLinf wrapper-only tensors, such as value heads used during training, are not included. |
|
|
