[Question] Released SFT and RL checkpoints appear identical to NaVILA checkpoints

[vumainguyen20-hue]

Released SFT and RL checkpoints appear identical to NaVILA checkpoints

Hi, thank you for releasing MobileVLA-R1.

We downloaded the official weight.zip.part-* archive and compared both released checkpoints against the corresponding NaVILA checkpoints:

• weight/sft vs. a8cheng/navila-siglip-llama3-8b-v1.5-pretrain
• weight/rl vs. a8cheng/navila-llama3-8b-8f

For weight/rl, the SHA256 checksums are identical for every model component:

Component	Matching SHA256
LLM shard 1	d4df84977665ab9be02f987a4eccd56abb0ac7273f12f6906f50c13bc7110371
LLM shard 2	42c210a04fc723675f858b383d546d2d5e0d4aa857d4fb321d3e57da241796f9
LLM shard 3	d53f6cab87861720f85f22af13e756cd7cccf569c2f8c096a7ec38d513be43e4
LLM shard 4	f35c5720effc40db734ac67b232e0b97389f802251d5d4d2480987f91f6dc937
Vision tower	1ff7d080a98a3f014b057dca741d7be358b700565493a15dfd82cb98903b0c4b
MM projector	0921455072f43bfe67637a6d69faa7248b830c15dc67a6e9798b50e2aed6d376

The released weight/sft checkpoint also has identical SHA256 checksums to the NaVILA pretraining checkpoint:

Component	Matching SHA256
LLM shard 1	0b9720ddd92eddd05daa33950a529ff615180d0a4f1c2a2c8c981a9ad211b127
LLM shard 2	8d4782b4a69ef03845159ce1a15e272aadaaf134dc138d68f616098e8531729c
LLM shard 3	3acdd690e65c24f42a24581b8467af98bd3ca357444580f8012aacd2bd607921
LLM shard 4	f47bdecf5731ea77ca4cf27a85db016d737302e0a96db40e0a6e27a756c18c0d
Vision tower	d769e3a32a6a9bac72d4d93b989e44491f71b50f02bfa14cd9187758d4a68ff1
MM projector	ca78440c0ccf657e5dca905075fc054d7b3e340038aae03d784f52b56d06856d

The corresponding config.json files are identical as well. In addition, the trainer_state.json released under weight/rl matches the NaVILA 8-frame checkpoint. Recursive comparisons found no differences in the model payloads.

Therefore, the current archive appears to contain:

• weight/sft: the NaVILA pretraining checkpoint
• weight/rl: the NaVILA 8-frame navigation checkpoint

Neither checkpoint appears to contain the released MobileVLA-R1-specific depth tower, point-cloud tower, modality bridges, or other parameters expected from the SFT/GRPO training described in the paper.

Could you please confirm whether these are the intended MobileVLA-R1 checkpoints? If the NaVILA checkpoints were uploaded by mistake, could you provide the correct MobileVLA-R1 SFT and GRPO-trained weights, including the depth and point-cloud components required to reproduce the reported results?

Thank you.