This is a jointly fine-tuned LingBot-VLA 2.0 checkpoint for all ten tasks in the standard LIBERO Goal suite. It was trained only on standard lerobot/libero demonstrations; no LIBERO-Pro observations, perturbations, or initial states were used for training.

Standard LIBERO results

The primary result uses the upstream standard LIBERO implementation and task prompts (--benchmark libero --suites libero_goal). Each task was evaluated for 10 closed-loop trials using the official initial states, seed 7, and the standard 300-step Goal episode limit.

Task	Description	Successes	Success rate
0	open the middle drawer of the cabinet	4/10	40%
1	put the bowl on the stove	9/10	90%
2	put the wine bottle on top of the cabinet	7/10	70%
3	open the top drawer and put the bowl inside	3/10	30%
4	put the bowl on top of the cabinet	9/10	90%
5	push the plate to the front of the stove	3/10	30%
6	put the cream cheese in the bowl	3/10	30%
7	turn on the stove	10/10	100%
8	put the bowl on the plate	6/10	60%
9	put the wine bottle on the rack	0/10	0%
Overall	LIBERO Goal	54/100	54%

The evaluation used prompt_source=task and reports 100 official-init episodes. Raw per-task JSON, logs, videos, and summary.json are included under evaluation/standard_libero_goal_10x10/.

Training and data contract

Base model: robbyant/lingbot-vla-v2-6b@11c703bf6a5c1f45b3b69168482da11fdbba53d7
Stage-two initialization: the corrected Goal task-7 checkpoint at local step 1400
Upstream source: robbyant/lingbot-vla-v2@951475ae1b1d87553e7dc47c97b53a3d695c0d13
Dataset: lerobot/libero@a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4
Training scope: all ten libero_goal tasks, 428 episodes / 52,042 frames
Stage-two optimizer updates: 2,000
Optimizer/LR: Muon, cosine 3e-5 to 3e-6, 2% warmup
Effective batch: 7 across physical GPUs 1–7 (7 × NVIDIA A100 80GB)
Action: 7D relative EEF delta (xyz + axis-angle + gripper)
State: 8D (eef xyz + axis-angle + gripper qpos(2))
Cameras: agent view as camera_top, eye-in-hand as camera_wrist
Action chunk: 50; official 55D unified head retained

The training pipeline uses the corrected absolute LeRobot sample index for filtered episode boundaries and excludes action_is_pad positions from the VLA loss mask.

Files

The root config/tokenizer files and six model-*.safetensors shards form the Hugging Face checkpoint. Reproducibility files include:

training/lingbotvla_cli.yaml: resolved model/training configuration
training/libero_goal_suite_fast_resolved.yaml: rendered run configuration
training/libero.yaml: robot/action/camera mapping
training/norm_stats.json: exact training and evaluation statistics
training/pins.json: pinned model, source, data, and teacher revisions
evaluation/standard_libero_goal_10x10/: standard LIBERO evidence
MANIFEST.json and SHA256SUMS: provenance and integrity metadata

Evaluation example

bash

python libero_eval/run_eval.py \
  --model Fisher-Wang/lingbot-vla-v2-6b-libero-goal \
  --commit-id <fixed-hugging-face-commit> \
  --backbone lingbot-vla-v2 \
  --benchmark libero \
  --suites libero_goal \
  --num-trials 10 \
  --gpus 0 \
  --lingbot-norm-stats training/norm_stats.json \
  --qwen3-vl-path /path/to/Qwen3-VL-4B-Instruct

The Qwen3-VL path is needed for processor/tokenizer construction; its weights are not included in this release.

Limitations

Task 9 failed all 10 recorded trials.
This release was jointly trained on LIBERO Goal only, not the standard Spatial, Object, LIBERO-10, or LIBERO-90 suites.
LIBERO-Pro runs are robustness diagnostics and are not the primary score for this standard-LIBERO-trained checkpoint.
The stage-two run covers about 27% of one shuffled data epoch by sample exposure; additional balanced training may improve the weaker tasks.