hiraeth111/uav-joint-baseline-weights

UAV joint baseline fine-tuning: initial checkpoints

Three initial UAV checkpoints prepared for the joint baseline fine-tuning branch. These are the released bases and their PyTorch conversions, before the planned joint fine-tuning, not new experimental results.

Model	Ready-to-use directory
OpenVLA-UAV	OpenVLA-UAV — complete released checkpoint, all four tensor shards and native tokenizer/processor/model code
IndoorUAV-Agent / π₀	IndoorUAV-pi0-pytorch — BF16 PyTorch conversion, with the publisher's existing UAV LoRA fused
HUGE π₀.₅-UAV	HUGE-pi05-pytorch — BF16 PyTorch conversion

The package includes the shared paligemma_tokenizer.model, model configs, conversion reports and original normalization assets for provenance. Joint fine-tuning uses fresh statistics from its own frozen training bundle. The two converted directories are already PyTorch checkpoints and need no additional JAX conversion.

Download

hf download hiraeth111/uav-joint-baseline-weights \
  --local-dir /data/uav-baseline-weights

cd /data/uav-baseline-weights/baseline-joint-ft-20261108
sha256sum -c SHA256SUMS

Copy the contents of baseline-joint-ft-20261108/ into the branch's checkpoints/, preserving the directory names, or specify the model directory with --checkpoint and shared tokenizer with --tokenizer.

See the package README, file manifest, training handoff, and conversion audit for exact conventions, environment requirements and numerical validation boundaries.

Sources and attribution

• OpenVLA-UAV, UAV-Flow code.
• IndoorUAV-Agent, published checkpoint 29999.
• HUGE_PI05, source revision a66531a8991aa585b8e2a356900d5fb07475fdd7; HUGE-Bench code.

Upstream weights retain their original authorship, licenses and access conditions. This collection does not relicense the models.