Kartik Nagpal
@kartiknagpal.bsky.social
PhD Student at University of California Berkeley | Multi-Agent Reinforcement Learning for Safety-Critical Robotics
kartik-nagpal.github.io
kartik-nagpal.github.io
💻 Additional details at our project page: iconlab.negarmehr.com/LLM-M…
📜 Read the full paper on Arxiv: arxiv.org/abs/2���
🖋 Authors: Kartik Nagpal, Dayi Dong, JB Bouvier, Negar Mehr
📜 Read the full paper on Arxiv: arxiv.org/abs/2���
🖋 Authors: Kartik Nagpal, Dayi Dong, JB Bouvier, Negar Mehr
https://iconlab.negarmehr.com/LLM-M…
March 8, 2025 at 10:13 PM
💻 Additional details at our project page: iconlab.negarmehr.com/LLM-M…
📜 Read the full paper on Arxiv: arxiv.org/abs/2���
🖋 Authors: Kartik Nagpal, Dayi Dong, JB Bouvier, Negar Mehr
📜 Read the full paper on Arxiv: arxiv.org/abs/2���
🖋 Authors: Kartik Nagpal, Dayi Dong, JB Bouvier, Negar Mehr
🏁 After training, our decentralized agent policies no longer need our LLM critic, and so have no slow down at runtime!
💪🏼 We also showcase an extension, LLM-TACA, which allows for explicit task assignment, producing even more improvements over existing methods!
💪🏼 We also showcase an extension, LLM-TACA, which allows for explicit task assignment, producing even more improvements over existing methods!
March 8, 2025 at 10:13 PM
🏁 After training, our decentralized agent policies no longer need our LLM critic, and so have no slow down at runtime!
💪🏼 We also showcase an extension, LLM-TACA, which allows for explicit task assignment, producing even more improvements over existing methods!
💪🏼 We also showcase an extension, LLM-TACA, which allows for explicit task assignment, producing even more improvements over existing methods!
⚙️ We train our decentralized agent policies directly on this feedback, enabling us to learn complex and powerful coordination behaviors!
🏆 LLM-MCA surpasses current MARL baselines across multiple common benchmarks! Including partially-observable scenarios!
🏆 LLM-MCA surpasses current MARL baselines across multiple common benchmarks! Including partially-observable scenarios!
March 8, 2025 at 10:13 PM
⚙️ We train our decentralized agent policies directly on this feedback, enabling us to learn complex and powerful coordination behaviors!
🏆 LLM-MCA surpasses current MARL baselines across multiple common benchmarks! Including partially-observable scenarios!
🏆 LLM-MCA surpasses current MARL baselines across multiple common benchmarks! Including partially-observable scenarios!
‼️In our AAMAS2025 paper "Leveraging Large Language Models for Effective and Explainable Multi-Agent Credit Assignment", we propose a centralized LLM credit critic we call LLM-MCA!
🎯 LLM-MCA translates sparse environment rewards to individualized numerical feedback for each agent!
🎯 LLM-MCA translates sparse environment rewards to individualized numerical feedback for each agent!
March 8, 2025 at 10:13 PM
‼️In our AAMAS2025 paper "Leveraging Large Language Models for Effective and Explainable Multi-Agent Credit Assignment", we propose a centralized LLM credit critic we call LLM-MCA!
🎯 LLM-MCA translates sparse environment rewards to individualized numerical feedback for each agent!
🎯 LLM-MCA translates sparse environment rewards to individualized numerical feedback for each agent!
As anyone who has done a group project can tell you, the effort toward finishing a task is rarely split evenly among the agents in the team, and as a result this "credit assignment problem" is very difficult.
March 8, 2025 at 10:13 PM
As anyone who has done a group project can tell you, the effort toward finishing a task is rarely split evenly among the agents in the team, and as a result this "credit assignment problem" is very difficult.
A recurring problem with our current multi-agent reinforcement learning (MARL) paradigms has been how we separate the contribution of each agent to achieving a reward.
March 8, 2025 at 10:13 PM
A recurring problem with our current multi-agent reinforcement learning (MARL) paradigms has been how we separate the contribution of each agent to achieving a reward.