SocialNav SAFE-GRPO checkpoints (mirror)

Inference files of the two released SocialNav checkpoints (Chen et al., "SocialNav: Training Human-Inspired Foundation Model for Socially-Aware Embodied Navigation", CVPR 2026, arXiv 2511.21135), mirrored for the arena-planners socialnav planner. The optimizer, scheduler, RNG and trainer state files of the source repositories are left out.

Source: ModelScope, https://www.modelscope.cn/models/zjugyn/SocialNav-Qwen2-VL-SAFE-GRPO (revision 2c729f68293db99f834c28ed8619ed27ee09a040) and https://www.modelscope.cn/models/zjugyn/SocialNav-Qwen2.5-VL-SAFE-GRPO (revision bf3368e84e8b22df4e002ca8e580c34b28c61be0). Code: https://github.com/AMAP-EAI/SocialNav

License: Apache-2.0, as stated in the ModelScope model cards. The Qwen2-VL checkpoint derives from Qwen/Qwen2-VL-2B-Instruct (Apache-2.0), the Qwen2.5-VL checkpoint from Qwen/Qwen2.5-VL-3B-Instruct (Qwen Research License Agreement, non-commercial research use), whose terms apply to the respective variant.

Files

File	sha256
SocialNav-Qwen2-VL-SAFE-GRPO/added_tokens.json	2351a6919c40fdaa47f99fc46dae8636103653ba6b4e70ad1946d45bbcf97c11
SocialNav-Qwen2-VL-SAFE-GRPO/chat_template.jinja	a0bc6f6fc7a29a80017a433e8f03a1cc1236e838a944a2d034295a60c4f2fddb
SocialNav-Qwen2-VL-SAFE-GRPO/config.json	0d2739c5c8e4025fb38522a306d26e5979e96272f34e6de0ee0fa534f349940c
SocialNav-Qwen2-VL-SAFE-GRPO/generation_config.json	0e2ceefb956c2283d64b1dae9ed23ee769898156b6c550db4ddb122ff0d67d5d
SocialNav-Qwen2-VL-SAFE-GRPO/merges.txt	8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5
SocialNav-Qwen2-VL-SAFE-GRPO/model.safetensors	56360bc5cc9b9f54d4fa9a2cc1c9c56ca017aebe52d7e615900ba3d12db22976
SocialNav-Qwen2-VL-SAFE-GRPO/preprocessor_config.json	29d27c703e5186231bc2ca07a876738a9f67ef5cb300ea2f72cf4048170d090d
SocialNav-Qwen2-VL-SAFE-GRPO/special_tokens_map.json	c2ed7c65623c393ca83edeca1aaac2f46533ac7e8c12eef934902e5d507e619f
SocialNav-Qwen2-VL-SAFE-GRPO/tokenizer.json	f1af02706374dba3446428a64e012321dc2413b78f7a867f75decef101d1ccaa
SocialNav-Qwen2-VL-SAFE-GRPO/tokenizer_config.json	37df7cf0d03541a61d15207e71dea17e72f5b89118055633efa27e8c6b3b56b3
SocialNav-Qwen2-VL-SAFE-GRPO/vocab.json	ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910
SocialNav-Qwen2.5-VL-SAFE-GRPO/added_tokens.json	9627ffe1755363c12c7cee128d1022f16d0e5b191c31cbae0d1f24dbc4883158
SocialNav-Qwen2.5-VL-SAFE-GRPO/chat_template.jinja	a0bc6f6fc7a29a80017a433e8f03a1cc1236e838a944a2d034295a60c4f2fddb
SocialNav-Qwen2.5-VL-SAFE-GRPO/config.json	fbd37fa85595aec3a4fe972733f575fdafc7b3e391499d175c121640172c77df
SocialNav-Qwen2.5-VL-SAFE-GRPO/generation_config.json	ab4c38cc7bd30c21bd70631dd67f8138830d06685f1858865b6b0436262e660c
SocialNav-Qwen2.5-VL-SAFE-GRPO/merges.txt	8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5
SocialNav-Qwen2.5-VL-SAFE-GRPO/model-00001-of-00002.safetensors	268ab090cdb68476c19d258409554bdce31afd6bd9454701467b3be4b2f1f237
SocialNav-Qwen2.5-VL-SAFE-GRPO/model-00002-of-00002.safetensors	3dd23ee3ce57534dd0c96bdfc6a8896ebb6e0b88b8df13ce2336d13a66a03d33
SocialNav-Qwen2.5-VL-SAFE-GRPO/model.safetensors.index.json	67a1845aee238beb3188917903c9b466733ffe026e44b1ac6606118868e247e5
SocialNav-Qwen2.5-VL-SAFE-GRPO/preprocessor_config.json	549c158011407dfb750d9ec578047cf76f5bfe365cd0aa069a50137d3f98d9dd
SocialNav-Qwen2.5-VL-SAFE-GRPO/special_tokens_map.json	c2ed7c65623c393ca83edeca1aaac2f46533ac7e8c12eef934902e5d507e619f
SocialNav-Qwen2.5-VL-SAFE-GRPO/tokenizer.json	5db06f405a3647e1ac28f333ef2e440fa18f818a3062deab109556971552f4a2
SocialNav-Qwen2.5-VL-SAFE-GRPO/tokenizer_config.json	2b93022238449ba45eef372bc7a58e1413be1c042f99d2815114c81f8b8c1d6b
SocialNav-Qwen2.5-VL-SAFE-GRPO/vocab.json	ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910

Citation

@article{chen2025socialnav,
  title={SocialNav: Training Human-Inspired Foundation Model for Socially-Aware Embodied Navigation},
  author={Ziyi Chen and Yingnan Guo and Zedong Chu and Minghua Luo and Yanfen Shen and Mingchao Sun and Junjun Hu and Shichao Xie and Kuan Yang and Pei Shi and Zhining Gu and Lu Liu and Honglin Han and Xiaolong Wu and Mu Xu and Yu Zhang},
  journal={arXiv preprint arXiv:2511.21135},
  year={2025}
}