longnav-objectnav-flow-nopose-cotrain-2p5hz-merged
Weights-merged twin of Aasdfip/longnav-objectnav-flow-nopose-cotrain-2p5hz.
Same policy, same step (trainer_state.json global_step 12000) — the LoRA is folded into
the base weights instead of shipped alongside them. Numerically identical: verified as
merged == base + (alpha/r)·B@A on a probe module, residual 1.5e-3 (bf16 rounding).
Why this exists
RL fine-tuning on top of this policy is run with the SFT LoRA already merged into the
base, so the RL run's saved adapter contains the RL delta alone — about 1–2% of the
SFT adapter's magnitude. Evaluating such an adapter against the plain adapter repo means
composing two LoRAs (rank-stacking A/B into an r=256/alpha=512 adapter), and getting it
wrong is quiet: apply the RL delta to the raw base and you evaluate the pretrained VLM
plus a 1% perturbation, which scores near zero and raises nothing.
With this repo, an RL adapter loads directly on top:
model = AutoModelForImageTextToText.from_pretrained("Aasdfip/longnav-objectnav-flow-nopose-cotrain-2p5hz-merged")
model = PeftModel.from_pretrained(model, "<rl-run>/checkpoints/checkpoint_XXX") # RL delta only
The adapter repo remains the canonical artifact: a merged model cannot be diffed against the base to recover the LoRA.
Files
| file | what |
|---|---|
model.safetensors, config.json | full merged weights (no adapter/) |
turn_vector_head.pt | flow-matching action head + pose encoder |
turn_vector_head_config.json | head config, including the <pose> modality spec |
trainer_state.json | training curve, for provenance |
| tokenizer / preprocessor files | as trained |
Optimizer and RNG state are not included — inference checkpoint, not a resume point.
Running it
Identical to the adapter repo, --ckpt pointed here. This policy is evaluated with no
pose input: it declares a <pose> modality and was trained with pose on half its ObjectNav
data, but scores better with the feed switched off entirely (0.663 vs 0.614 oracle success
on sample101, n=101).
python scripts/eval_objectnav_policy.py \
--episodes <hm3d objectnav val split> \
--scene-root <scene_datasets> \
--output-dir <out> \
--episode-ids "$(cat sample101_ids.txt)" \
--ckpt <this checkpoint> \
--policy-python <env with transformers> \
--policy-sys-path <longnav src> \
--policy-backend flow_rollout --policy-seed 1000 \
--max-steps 175 --dt 0.04 --gap 10 \
--success-distance 1.0 --distance-to VIEW_POINTS \
--seed 0 --navmesh dataset --fresh-sim-per-episode \
--auto-stop --auto-stop-delay 15 \
--no-pose-injection \
--pid-preset baseline --keep-tracking-series
--no-pose-injection is required for the pose-free arm and is not in a public release of
the harness at the time of upload; without it there is no way to remove the modality
(feeding zeros is not equivalent — it is an in-distribution-looking input the model never
trained on). The simulator and the model cannot share a Python environment, which is what
--policy-python / --policy-sys-path are for.