Panpan Cai

This is Vec-POMDP: ✅ A CPU-native parallel POMDP planner ✅ Multi-threading + global vectorization + local vectorization + SIMD ✅ 227×–1073× speedup Website: sii-boluomonster.github.io/VecQMDP-websit… Paper: arxiv.org/abs/2602.08334 Code: github.com/SII-BoluoMonst…

POMDPs are great for tackling uncertainties—hidden states, stochastic action outcomes, noisy observations. But they don’t scale. We boost POMDP planning with vectorization + SIMD. In urban driving, we get: ☑️ 9 ms planning time ☑️ SOTA on nuPlan ☑️ no training

✅ Planning framework for open-ended uncertainty ✅ LLM-generated belief + Bayesian filtering + POMDP planning ✅ 85% success on very ambiguous planning tasks with many hidden objects Website：tru-pomdp.github.io Paper：arxiv.org/pdf/2506.02860 Code：github.com/RoboticSJTU/tr…

In an open world, robots face both uncertainty and unknowns. While POMDPs handle the former elegantly, they break on the latter, where no clear hypotheses exist. So, let's bridge POMDPs with LLMs — bringing structured reasoning to open-world robotics!

✅ Disentangled intent & execution tokens ✅ Intent-to-execution reasoning ✅ One-shot transfer via intent token injection ✅ 12.2% stronger generalization Website: renming-huang.github.io/MINT Code: github.com/RenMing-Huang/… Model: huggingface.co/huangrm/MINT-l…

The biggest bottleneck for using VLAs in the real world? - Generalization! But it’s not just about changing layouts, lighting, ... It’s: ☑️ Compositionality: A,B,C → combos? ☑️ Transferability: learn a new task from 1 demo? That’s exactly what MINT offers.

This is UniDomain: ☑️ 3,137 operators, 2,875 predicates, and 16,481 causal edges ☑️ Pretrain once, then reuse for real-world task planning ☑️ +58% success, +160% optimality Website: roboticsjtu.github.io/UniDomain/ Code: github.com/RoboticSJTU/Un… Data: huggingface.co/datasets/SII-P…

Can LLMs alone nail robot planning? No. ❌ Physical constraints ❌ Long-horizon planning (success quickly becomes 0%) ❌ Optimality What if we learn from real-world demonstrations, and turn that into a unified symbolic world model (PDDL)?

POMDPs are great for tackling uncertainties—hidden states, stochastic action outcomes, noisy observations... But they don’t scale. We boost POMDP planning with vectorization + SIMD. In urban driving, we get: - 9 ms planning time! - 227×–1073× speedup! youtu.be/FONnp1gvLqE

In an open world, robots face both uncertainty and unknowns. While POMDPs handle the former elegantly, they break on the latter, where no clear hypotheses exist. So, let's bridge POMDPs with LLMs — bringing structured reasoning to open-world robotics! youtu.be/_UCBoBJ5YxI

The biggest bottleneck for using VLA in the real world? Generalization! But it’s not just about changing layouts, lighting, ... It’s: - compositionality: A,B,C → combos? - transferability: learn a new task from 1 demo? That’s exactly what MINT offers. youtu.be/p2XRFBYGh_A

Can LLMs alone nail robot planning? No. ❌ Physical constraints ❌ Long-horizon planning (success quickly becomes 0%) ❌ Optimality What if we learn from real-world demonstrations, and turn that into a unified symbolic world model (PDDL)? youtu.be/GHFB_F_HLA0

An ultimate challenge in autonomous driving is to interact with human crowds. We provide a simulator to facilitate tackling the challenge. Check out our #ICRA2020 paper "SUMMIT: A Simulator for Urban Driving in Massive Mixed Traffic". Youtube video: youtu.be/rF-n52gnMw0