PhD Student @UTAustin | MS @GeorgiaTech | Former Intern @Microsoft @NASAJPL @IBMResearch
Austin
Joined September 2025
- Tweets10
- Following35
- Followers13
- Likes2
Pinned Tweet
[1/8] New social navigation paper + benchmark: SocialNav-SUB 🚶🤖 Recent work puts VLMs on robots for navigation, but can they really interpret scenes and extract key details for social navigation? 🔎 larg.github.io/socialnav-sub
Michael Munje retweeted
How can robots follow complex instructions in dynamic environments?
🤖 Meet ComposableNav — a diffusion-based planner that enables robots to generate novel navigation behaviors that satisfy diverse instruction specifications on the fly — no retraining needed.
📄 Just accepted to CoRL 2025
🔗 Project: amrl.cs.utexas.edu/Composabl…
A Thread (1/8)
A huge thanks to my collaborators for making this work possible! @ChenTangMark @dafeijing @ZichaoHu99 @yifengzhu_ut @cuijiaxun @GarrettWarnell @Joydeepb_robots @PeterStone_TX
[8/8] 🤝 SocialNav-SUB is a human-grounded check on whether VLMs understand social navigation scenes ✨ Please read our paper for more info: arxiv.org/abs/2509.08757 #Robotics #VLM #SocialNavigation
[7/8] SocialNavSUB is also fully open-source, actively maintained, and easily extendable to customized prompts and/or additional VLMs! Pull requests are always welcome! github.com/LARG/SocialNavSUB
[6/8] 🧪 Does chain-of-thought (using spatial/spatiotemporal VQAs first) improve social reasoning? ✅Yes. Does BEV context help models? ⚖️ Model-dependent (sometimes a lot). Does better spatial(temporal) context improve social reasoning? ✅Yes.
[5/8] 📊 Do today’s VLMs agree with human judgments? We find that they still trail behind humans and simple rule-based baselines.
[4/8] 👥 We collected human data from an IRB-approved human-subject study to construct our benchmark and evaluate whether models align with human judgments in social navigation scenes.
[3/8] SocialNav-SUB features real-world social navigation scenarios built from SCAND scenarios @ 4 Hz → PHALP tracking → front-view & BEV with labeled pedestrians, combining them with a set of carefully designed questions to create our VQA prompts (5k in total).
