lingbot-vla-lora_tuned_1500

LoRA fine-tuning of LingBot-VLA 2.0 (6B) on the RoboTwin 2.0 (Aloha-AgileX) robotic manipulation benchmark.

Code, evaluation results and reproduction reports: https://github.com/wuuwuu26/lingbot-vla-lora_tuned_1500

License: this repository inherits the terms of its base model robbyant/lingbot-vla-v2-6b, which does not declare a license on the Hub. Please refer to the upstream project for usage terms.

Model details

• Base model: robbyant/lingbot-vla-v2-6b
• Training progression:
  1. Full-parameter SFT, 11000 steps (official RoboTwin recipe, no hyper-parameter changes)
  2. Continued SFT to 6000 steps (lora6000_base, lr 1e-4)
  3. LoRA 1500 steps on that base — the submitted checkpoint (lora_tuned_1500)
• LoRA configuration: rank 64 / alpha 128, scope action_expert, target modules q_proj,k_proj,v_proj,o_proj, lr 5e-5, merged back into the base
• Training data: the 50 clean tasks of RoboTwin 2.0 (Aloha-AgileX), 50 episodes each. No randomized data and no external data were used.
• Action space: 55-dimensional unified representation (14 arm joints + 14 end-effector poses + 2 grippers + …)

Files

File	Size	SHA-256
model-00001-of-00003.safetensors	2.0 GB	7bff0166a4cc6e5f66d2b52b144f863679c81ef890c719e1ced5c499cfe22cc4
model-00002-of-00003.safetensors	5.0 GB	2ba12d8f780fa0df533d764e01b0a09528b1901e02c79c23af00ca96c5d89267
model-00003-of-00003.safetensors	4.6 GB	495150c7f55d83f65554580179bffa2a33f078ce9d14073f60a57f3ee7957aac

The repository also contains config.json, model.safetensors.index.json, lingbotvla_cli.yaml (the training config) and the tokenizer / preprocessor files, so downloading this repository yields a complete checkpoint.

Download

hf download Dwyane/lingbot-vla-lora_tuned_1500 --local-dir ./checkpoint

Behind a slow or restricted connection, use the mirror:

export HF_ENDPOINT=https://hf-mirror.com
export HF_HUB_DISABLE_XET=1
hf download Dwyane/lingbot-vla-lora_tuned_1500 --local-dir ./checkpoint

Verify the download:

cd checkpoint && sha256sum model-*.safetensors   # compare with the table above

Evaluation results

Planner cuRobo v0.7.8, use_length=25, bf16, 10 episodes per task.

Setting	Tasks	Score	Success rate
clean2clean (seen)	50 / 50	314 / 500	62.8%
clean2random (unseen)	50 / 50	185 / 500	37.0%

For comparison, the earlier 11000-step full-parameter SFT checkpoint scored 48.55% (670/1380) on clean and 24.00% (180/750) on randomized.

Self-evaluation used cuRobo. If official re-evaluation falls back to MPLib (the platform image has no cuRobo), absolute success rates for the same weights will be lower — a tooling difference, not a policy difference. Official re-evaluation takes precedence.

Per-task details and the full tuning notes are in the GitHub repository (results/ and docs/).