| --- |
| license: apache-2.0 |
| library_name: rlinf |
| pipeline_tag: robotics |
| tags: |
| - reinforcement-learning |
| - robotics |
| - embodied-ai |
| - lingbotvla |
| - robotwin |
| - grpo |
| --- |
| |
| # RLinf LingbotVLA Click Bell GRPO |
|
|
| This repository contains an RLinf checkpoint for LingbotVLA GRPO fine-tuning on the RoboTwin `click_bell` task. |
|
|
| ## Checkpoint Format |
|
|
| The checkpoint is provided in RLinf actor checkpoint format: |
|
|
| ```text |
| actor/model_state_dict/full_weights.pt |
| actor/dcp_checkpoint/.metadata |
| actor/dcp_checkpoint/__5_0.distcp |
| actor/dcp_checkpoint/__7_0.distcp |
| ``` |
|
|
| For evaluation in RLinf, load the model through `runner.ckpt_path`: |
|
|
| ```bash |
| runner.ckpt_path=/path/to/actor/model_state_dict/full_weights.pt |
| ``` |
|
|
| Use the LingbotVLA RoboTwin SFT base configuration from: |
|
|
| ```text |
| robbyant/lingbot-vla-4b-posttrain-robotwin |
| revision: 3e0c7c476bde3daaac00f79f3741a292a299f60a |
| ``` |
|
|
| ## Evaluation |
|
|
| Latest local regression evaluation on RoboTwin `click_bell` random setting: |
|
|
| | Checkpoint | Task | Setting | Trajectories | Max Steps | eval/success_once | eval/return | |
| | --- | --- | --- | ---: | ---: | ---: | ---: | |
| | RLinf-lingbotvla-click-bell-grpo | click_bell | random | 320 | 400 | 0.9875 | 6.85 | |
|
|
| Evaluation settings: |
|
|
| ```text |
| config: robotwin_click_bell_grpo_lingbotvla_eval |
| algorithm.eval_rollout_epoch=1 |
| algorithm.sampling_params.temperature_eval=-1 |
| env.eval.total_num_envs=320 |
| env.eval.max_episode_steps=400 |
| env.eval.max_steps_per_rollout_epoch=400 |
| env.eval.use_fixed_reset_state_ids=False |
| env.eval.seeds_path=null |
| env.eval.video_cfg.save_video=False |
| ``` |
|
|
| The evaluation logs are kept locally under: |
|
|
| ```text |
| /mnt/public/lwb/artifacts/lingbot-vla-eval/click_bell_regression/20260615_140703 |
| ``` |
|
|
