This is a compact vision-language-action policy built on Qwen/Qwen3.5-0.8B and supervised-fine-tuned on the combined LIBERO training data. It uses a parallel categorical OFT action head and is packaged as a portable checkpoint for the QwenDiscreteOFT integration in SimpleVLA-RL.

The checkpoint contains the complete VLM and action-head weights. In particular, base_vlm in qwen_discrete_oft.json is provenance metadata; the SimpleVLA-RL runtime does not download or import vla-mono, StarVLA, or OpenVLA-OFT.

Model details

Field	Value
Parameters	867,163,968
Visual input	Agent view + wrist view, 224×224 RGB
Action chunk	8 steps × 7 dimensions
Discretization	256 bins over [-1, 1]
Gripper	Binary categorical dimension
Normalization	Dataset min/max
SFT data	`libero_all`
SFT steps	30,000
Seed	42

LIBERO evaluation

Evaluation used deterministic argmax actions, the two-view inference pipeline, 50 initial states per task, and all 10 tasks in each suite. There were 2,000 episodes in total and no environment/infrastructure errors.

Suite	Successes	Episodes	Success rate
LIBERO-Spatial	483	500	96.6%
LIBERO-Object	496	500	99.2%
LIBERO-Goal	472	500	94.4%
LIBERO-10	442	500	88.4%
Overall	1,893	2,000	94.65%

The episode horizons were 220, 280, 300, and 520 environment steps for Spatial, Object, Goal, and LIBERO-10 respectively. See eval_results.json for machine-readable counts and protocol metadata.

SimpleVLA-RL usage

Use the QwenDiscreteOFT branch of JustinKai0527/VLASafety:

bash

cd SimpleVLA-RL
pip install -r requirements-qwen35.txt

SFT_MODEL_PATH=jasper0314/starvla-libero-qwen3.5-0.8b-discrete-oft \
CKPT_PATH=/path/to/rl-checkpoints \
bash examples/run_qwen35_discrete_oft_rl_libero.sh

Use a dedicated environment for this model because Qwen3.5 requires Transformers 5.3, while the repository's legacy OpenVLA path uses an older Transformers version. The RL algorithm and trainer remain SimpleVLA-RL's existing veRL/GRPO-PPO path; the integration supplies model loading, multimodal prompting, discrete rollout responses, log-probability recomputation, and portable checkpoint saving.

This is a robotics policy, not a text-generation checkpoint and not a drop-in AutoModel model. Use it only in a compatible simulator/runtime and validate actions and safety constraints before any real-robot deployment.