Ruiyang Wang

PhD Student in Electrical and Computer Engineering, Duke University

I am a third-year PhD student in Electrical and Computer Engineering at Duke University, advised by Prof. Miroslav Pajic in the Cyber-Physical Systems Lab. My research makes multi-robot teams work together reliably, combining foundation models, graph learning, and neuro-symbolic reasoning for coordination and task allocation.

Now I build harness systems that verify and repair the action plans LLMs and VLMs write, so robots complete long-horizon tasks and satisfy safety constraints. Before Duke, I earned my M.S. in Robotics and B.S. in Mechanical Engineering at the University of Michigan, working on safety-critical control with model predictive control (MPC) and control barrier functions (CBFs).

Research interests: Multi-Robot Systems · LLM/VLM Planning for Robotics · Harness Systems for Safe Long-Horizon Planning · Neuro-Symbolic AI

Seeking Summer 2027 research internships on harness systems for LLM/VLM robot planners, long-horizon task planning, and multi-robot systems.

Email CV Google Scholar GitHub LinkedIn
Ruiyang Wang

News

Research & Publications

GAVEL example: a robot in an apartment with seen and uncertain object locations, reasoning about which task to do first

GAVEL: Graph World Models for Verified and Efficient Long-Horizon LLM Task Planning

Ruiyang Wang, Hao-Lun Hsu, Swarajh Mehta, Jiwoo Kim, Zhihao Dou, Miroslav Pajic

Submitted to ICRA 2027

LLM-generated robot plans often skip prerequisite steps or break embodiment constraints. GAVEL gives the LLM an explicit graph world model: it rolls each plan forward before execution, directly repairs failures whose fix follows from the action model, and sends only failures that need semantic reasoning back to the LLM. For multi-task instructions, it keeps beliefs over where unseen objects are and reorders the remaining tasks as the robot observes more.

  • 41.2% → 91.8% success on 100 long-horizon BEHAVIOR-1K tasks (Qwen3-8B)
  • 19.9% → 92.6% success on 500 multi-task instructions under partial observability
  • A 4B local model with GAVEL (75.2%) outperforms frontier hosted LLMs used alone (at most 38.6%)
LLM Task Planning World Models Plan Verification
Paper Video Featured by DAIR.AI
Three robots report conflicting observations of a cardboard box: which robots should be trusted, and does the box exist?

Who to Trust in Multi-Robot Systems: Joint Robot and Object Hypothesis Trust Inference via Neural-Symbolic Cross-Validation

Ruiyang Wang, Hao-Lun Hsu, Haocheng Meng, Songqiao Xie, Miroslav Pajic

Submitted to RA-L

When robots share what they see, a compromised teammate can inject fake objects or hide real ones. NeST-Bayes infers which robots, and which reported objects, to trust using only whether robots' reports agree with each other: no prior trust, reputation, or side-channel signals. Each robot builds an ego-centric graph of agreements and contradictions, a heterogeneous graph attention network turns it into evidence, and Beta-distribution updates accumulate that evidence over time.

  • 98.8% / 98.3% robot / object precision against aggressive, optimized attackers
  • 92.5% / 97.3% robot / object precision in an occlusion-rich Webots warehouse with simulated Spot robots, without retraining
  • Highest robot precision even when 40–50% of the team is adversarial (89.1% vs. at most 73.8%)
Multi-Robot Trust Neuro-Symbolic Graph Neural Networks
Paper Video
A Unitree Go2 quadruped and an X500 drone given a language hint to search for a target at the far end of the main corridor

COMRES-VLM: Coordinated Multi-Robot Exploration and Search using Vision Language Models

Ruiyang Wang, Hao-Lun Hsu, David Hunt, Jiwoo Kim, Shaocheng Luo, Miroslav Pajic

Submitted to RA-L

A vision-language model acts as the team leader. Every 100 seconds it receives the merged team map, with the unexplored space cut into sensor-sized pieces behind each opening, and assigns every robot a piece, using a natural-language hint about the target when one is given. Between plans, each robot picks frontiers on its own while favoring its piece, so a late or missing plan never stalls the team.

  • Cuts search time by 25% versus Hungarian assignment with priors, across 40 maps including planning delays
  • Cuts search time by 33% versus Co-NavGPT, an LLM-based multi-robot navigation method
  • Evaluated in real-world experiments with two Unitree Go2 robots
VLM Coordination Multi-Robot Exploration Language-Guided Search
Paper
SAGR framework: a semantic reasoning layer where an LLM assigns robots to rooms from a semantic area graph, a frontier selection layer using recursive Hungarian assignment and TSP, and a local execution layer on each robot

Semantic Area Graph Reasoning for Multi-Robot Language-Guided Search

Ruiyang Wang, Hao-Lun Hsu, Jiwoo Kim, Miroslav Pajic

arXiv preprint

Instead of prompting a language model with raw maps or images, SAGR compresses the team's semantic occupancy map into a compact graph of rooms, where each node records its room type, frontiers, robots, and neighbors. An LLM reads this graph and the task (e.g. "find my jacket in the bedroom") and assigns robots to rooms; Hungarian matching and TSP ordering then handle frontier-level execution inside each room.

  • Up to 18.8% faster semantic search than the best baseline in large HM3D apartments
  • Fastest search at every environment scale while staying competitive on pure exploration
  • Compact prompts: under 500 tokens and about 2.5 s per coordination query
LLM Coordination Multi-Robot Search Scene Graphs
Paper
Neural Logic Machine pipeline: unary and binary predicate values pass through expand/reduce MLP layers to produce a score for each robot

Neuro-Symbolic Deadlock Resolution in Multi-Robot Systems

Ruiyang Wang, Bowen He, Miroslav Pajic

L4DC 2025

Decentralized robots squeezing through a narrow passage can freeze in a deadlock that no single controller escapes. NSDR resolves deadlocks after they occur: an Active-Passive paradigm lets selected robots yield without creating new persistent deadlocks, and a Neural Logic Machine learns which robot should yield. Because the learned logic rules do not depend on team size, a model trained on small deadlocks generalizes to larger teams.

  • Trained on 2–5-robot deadlocks; generalizes to 8- and 10-robot and unseen asymmetric deadlocks
  • With 10 robots: 23% lower total arrival time than greedy and 12% lower than adaptive repulsive force
  • Resolves two-way traffic in which the repulsive-force baseline deadlocks again
Neuro-Symbolic Deadlock Resolution Distributed MPC
Paper Slides

Algorithms for Finding Compatible Constraints in Receding-Horizon Control of Dynamical Systems

Hardik Parwana*, Ruiyang Wang*, Dimitra Panagou (*equal contribution)

ACC 2024

When soft constraints, such as timed waypoints, conflict with hard constraints like obstacle avoidance, a receding-horizon controller becomes infeasible. Relaxing them with slack variables can compromise the hard constraints over time, so we drop incompatible soft constraints instead, choosing them with polynomial-time heuristics based on Lagrange multipliers.

  • Matches exhaustive search (12 of 14 waypoints) in 6.95 s vs. 3.61 h
  • Outperforms the slack-based Chinneck heuristic at lower computation time
  • Keeps safety and on-time goal arrival where slack relaxation fails
Receding-Horizon Control Control Barrier Functions Safety-Critical Control
Paper Code

Experience & Education

Research Experience

Ph.D. Researcher

August 2024 - Present

Cyber-Physical Systems Lab, Duke University

Built VLM coordination, graph-attention trust inference and neuro-symbolic deadlock resolution for multi-robot teams; now building harness systems that verify and repair LLM/VLM plans for long-horizon, safety-constrained robot tasks.

Graduate Research Assistant

December 2022 - December 2023

Distributed Aerospace and Control Lab, University of Michigan

Worked on constraint selection in receding-horizon control systems under Prof. Dimitra Panagou (ACC 2024).

Education

PhD in Electrical and Computer Engineering

August 2024 - April 2028 (Expected)

Duke University

Advisor: Prof. Miroslav Pajic

Master of Science in Robotics

August 2022 - December 2023

University of Michigan, Ann Arbor

Bachelor of Science in Mechanical Engineering

August 2018 - April 2022

University of Michigan, Ann Arbor

CV (PDF)