Qwen3-VL-4B OFT for Bridge + RT-1 (5K)
This repository contains a StarVLA QwenOFT WidowX policy trained on the
Bridge and Fractal/RT-1 LeRobot datasets. Only a 5K checkpoint is released.
[!WARNING]
The maintainer states in
starVLA issue #424
that the earlier uploaded checkpoint was incorrect, the original files were
removed, and a replacement is being retrained. Treat the current weights as a
research artifact pending a corrected release, not as a checkpoint expected
to reproduce the advertised 42.7% result.
Checkpoint identity
| Item | Value |
|---|
| Released file | checkpoints/steps_5000_pytorch_model.pt |
| Training step | 5,000 |
| Hub revision checked | c3fc8f028429ba14819bf3b16e098776b670c889 |
| File size | 9,785,060,316 bytes |
| SHA-256 / LFS object ID | 371cb744227687bb99bcad7f9ff2250cf06da75631359ad3eba4c6bc52570607 |
summary.jsonl describes later run activity, but it does not make a later
checkpoint available. This card does not relabel the artifact as a 35K or 100K
release.
Model and control contract
| Item | Value |
|---|
| Framework | StarVLA QwenOFT |
| Base VLM | Qwen3-VL-4B-Instruct |
| Action head | Two-block residual MLP, 2,560 input / 5,120 hidden / 7 output; direct L1 regression |
| Observation | Language + one image_0 RGB view at 224 x 224 |
| Robot state | Not sent by the matching SimplerEnv adapter |
| Action chunk | 16 x 7 |
| Action representation | delta end-effector action (delta_ee) |
| SimplerEnv normalization | oxe_bridge for WidowX Bridge evaluation |
| Action ensemble horizon | 7 |
Stale diffusion-related fields in the resolved run configuration do not turn
this QwenOFT artifact into a flow-matching policy; the selected framework and
head are regression-based.
Training data and settings
The bridge_rt_1 mixture combines oxe_bridge (1,284,318 transitions,
53,192 trajectories) and oxe_rt1 (3,449,678 transitions, 87,212
trajectories).
| Setting | Value |
|---|
| Per-device VLA batch size | 16 |
| Gradient accumulation | 1 |
| Optimizer | AdamW, betas (0.9, 0.95), epsilon 1e-8, weight decay 1e-8 |
| Warmup configured | 10,000 steps |
freeze_modules | Packaged boolean true; the public trainer expects module paths as a string, so this value names/selects no modules |
| Seed | 42 |
| Configured run target | 100,000 steps; released checkpoint is 5,000 |
| Training GPU count | Missing from the public artifact |
Evaluation: packaged historical logs
This repository contains 16 historical evaluation logs named for a 5K
checkpoint. Every log's final invocation identifies a file named
steps_5000_pytorch_model.pt and evaluates episodes [0, 24). The complete
set covers four tasks x four runs x 24 episodes = 384 episodes.
| Task | Successes / episodes | Success |
|---|
| Put carrot on plate | 49 / 96 | 51.04% |
| Put eggplant in basket | 85 / 96 | 88.54% |
| Put spoon on tablecloth | 30 / 96 | 31.25% |
| Stack green cube on yellow cube | 0 / 96 | 0.00% |
| Micro-average | 164 / 384 | 42.71% |
The per-run decimals sum to 164 successful episodes, so 42.71% is a
recomputation of the packaged logs only. The logs contain a filename but no
SHA-256/LFS object ID and therefore are not cryptographically bound to the
current checkpoint bytes. In light of the maintainer-confirmed upload mismatch,
this table must not be presented as the expected performance of the current
release. It is also not evidence for an unreleased longer-run checkpoint.
Known checkpoint mismatch
Issue #424 documents an independent reproduction gap and the maintainer's
subsequent confirmation that an incorrect checkpoint had been uploaded. The
issue remains open at the time of this card audit. A corrected artifact, its
immutable hash, and evaluation logs generated from that same hash are still
missing. User-reported reproduction numbers are useful diagnostic evidence but
are not promoted here as official checkpoint results.
Download and load
1hf download StarVLA/Qwen3VL-OFT-Bridge-RT-1 \
2 --local-dir playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
3
4export CKPT=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
5python deployment/model_server/server_policy.py \
6 --ckpt_path "$CKPT" \
7 --config_override framework.qwenvl.base_vlm=Qwen/Qwen3-VL-4B-Instruct \
8 --port 57700 \
9 --use_bf16
Use the WidowX Bridge adapter with oxe_bridge, delta-EEF actions, no state,
and the matching action ensemble. Keep the configuration and dataset statistics
beside the checkpoints/ directory.
Intended use and limitations
This early 5K checkpoint is retained for diagnosis and SimplerEnv research
while a corrected release is pending; it is not suitable for safety-critical
control. Expected performance of the current bytes is unknown, and performance
on physical robots or the RT-1 embodiment is unverified. Use only a trusted
.pt revision with the matching StarVLA code.
License status
This target repository did not previously publish a Model Card or a separate
LICENSE file. The checkpoint's weight license therefore needs maintainer
confirmation; the Qwen3-VL base-model and training-dataset terms still apply.
