Alexey Skrynnik

@FlexAppeall @Karolis_Ram Asynchronous PPO for large-scale experiments, i.e. github.com/alex-petrenko/…

@Tviskaron @Karolis_Ram By APPO you mean Augmented PPO? Like this? ojs.aaai.org/index.php/AAAI…

Sad to see that the update made the huge simplification of MineRL even less transparent. Without it, DeepMind’s Dreamer would get ~0 score. Before:”We following prior work and increase the speed at which blocks break” Now:”We follow the block breaking setting of prior work” 1/8

@_kei18 Glad to hear the honey is safe!😊

@Tviskaron No no, 🍯 survived ! Just forgot to include it in the photo. I enjoy lovely gifts l

開けても開けても出てくるマトリョーシカ🪆

@MikhailBurtsev 😄😄😄

When senior colleagues join the team at academic race.

What a way to wrap up @IgluContest! Our paper “IDAT: A Multi-Modal Dataset and Toolkit for Building and Evaluating Interactive Task-Solving Agents” accepted to @SIGIRConf including: 1) rich multi-modal dataset 2) A data collection tool 3) An online eval framework #SIGIR2025

Decentralized MAPF? :)

I’m happy to share that our paper, POGEMA: A Benchmark Platform for Cooperative Multi-Agent Pathfinding, has been accepted to the ICLR 2025 Conference! arxiv.org/abs/2407.14931 openreview.net/forum?id=6VgwE… See you in Singapore!

I’m happy to announce that our paper, MAPF-GPT: Imitation Learning for Multi-Agent Pathfinding at Scale, has been accepted to the AAAI 2025 Conference! arxiv.org/abs/2409.00134 github.com/CognitiveAISys…

@_kei18 Thank you, Keisuke :) However, I believe MAPF-GPT can be further improved by RL fine-tuning. P.S. LaCAM is great!

MAPF-GPT: Imitation Learning for Multi-Agent Pathfinding at Scale arxiv.org/abs/2409.00134 I like this direction, rather than developing better RL policies for MAPF.

Here are the links to the preprint: arxiv.org/abs/2409.00134 and the source code: github.com/Cognitive-AI-S…. The repository includes training code, pre-trained weights for the 2M, 6M, and 85M models (uploaded to HuggingFace), and a dataset of 1 billion observation/action pairs.

MAPF-GPT performs exceptionally well on unseen instances and outperforms state-of-the-art learnable solvers such as SCRIMP and DCC, having better runtime efficiency. (2/3)

I’m excited to announce our recent preprint titled MAPF-GPT, a GPT-like model designed for MAPF problems. It is trained using pure imitation learning on trajectories generated by LaCAM. (1/3)

PRL (Workshop) @ ICAPS 2024 *CALL* - If you're bridging the gap between AI Planning and Reinforcement Learning, join the PRL workshop - prl-theworkshop.github.io/prl2024-icaps/. Submissions due March 22

More details: ar5iv.labs.arxiv.org/html/2310.01207. Big thanks to my co-authors Anton Andreychuk, Maria Nesterova, Konstantin Yakovlev, and Aleksandr Panov. The second paper, which I will detail after the camera-ready version, focuses on using Neural MCTS for the LMAPF tasks. (7/7)

Against centralized MAPF algorithms, Follower excels, especially under a strict 1-second time constraint in scenarios with a large number of agents (over 160 for the Warehouse map). (6/7)

I'm excited to announce that two of my papers have been accepted at AAAI-2024, delving into lifelong multi-agent pathfinding (LMAPF) using RL and planning techniques. Here's the thread: (1/7).