CombatVLA - Efficient VLA Model by Amoy Group

What is CombatVLA?

CombatVLA is an innovative 3D action role-playing game (ARPG) specific model introduced by the Future Life Lab team of the Amoy Group.CombatVLA is a visual-linguistic-action (VLA) model built on a 3B parameter scale, which collects human player's game operation data through a motion tracker and transforms it into an "Action Thinking " (AoT) sequences for training. The model uses a three-stage incremental learning approach, from video-level to frame-level to truncation strategy, to significantly improve inference speed. In combat tasks, CombatVLA is able to make efficient decisions in real-time, such as avoiding attacks and launching skills, and its combat understanding is significantly better than other models.

Features of CombatVLA

• Efficient decision-making capacity: The ability to make precise decisions in real time in the complex combat environment of a 3D game, including actions such as dodging, attacking and recovering, with decision-making speeds far exceeding those of traditional models.
• Deep Combat Understanding: The model is able to deeply analyze the enemy's state and attack intent, reasoning out the best combat moves, with combat comprehension significantly better than other similar models.
• Instruction generation and execution: Support for generating specific keystroke commands to directly control the game character in battle, enabling efficient and precise execution of actions.
• Powerful generalization performance: Facing multiple game difficulties or brand new game scenarios, you can excel in combat tasks with your strong generalization ability.

CombatVLA's core strengths

• Efficient decision-making: Make quick combat decisions in complex 3D environments, with decision-making speeds up to 50 times faster than traditional models, improving game smoothness.
• deep understanding: Supports accurate assessment of enemy states and reasoning of optimal combat maneuvers, significantly outperforming other models and performing like a seasoned player.
• Strong generalization ability: Demonstrate great flexibility by adapting quickly and performing efficiently in different difficulties and new game scenarios.
• Precise command generation: The ability to directly output keystroke commands for precise control of the game character allows the player to focus on strategy rather than operational details.
• Advanced training techniques: Using three-stage incremental learning and adaptive action weight loss to ensure efficient model inference and accurate output of key actions.
• multi-scenario application: Applicable to multiple fields such as ARPG, game testing, e-sports training, content creation and robot control, with high practical value.

What is CombatVLA's official website?

• Project website:: https://combatvla.github.io/
• GitHub repository:: https://github.com/ChenVoid/CombatVLA
• arXiv Technical Paper:: https://arxiv.org/pdf/2503.09527

Who CombatVLA is for

• game developer: Help game developers test and optimize the combat system, quickly identify potential problems and make improvements to enhance the overall quality of the game.
• e-Sports player: Provide smart opponents for e-campaigners to help users practice combat skills and tactics, improve their competitive level, and better cope with various situations in real matches.
• gamer: Provide smarter AI opponents or teammates for regular players to increase the fun and challenge of the game and enhance the gaming experience.
• Game content creators: Assist creators in generating battle scenes and storylines, quickly build complex game levels and missions, and improve the efficiency and quality of content creation.
• Artificial intelligence researchers: As an advanced VLA model, it provides a platform for researchers to study multimodal interaction and reinforcement learning, and promotes the development of related technologies.