Motor-Skill-Learning | Takato Horii | Osaka University

Autonomous Motion Learning via Large Language Models

Mon, 01 Jan 2024 00:00:00 +0000

Using large language models (LLMs) and vision-language models (VLMs), the robot autonomously generates demonstrations from task instructions, enabling motion-skill learning without expert teaching or pre-defined motion primitives. Through an affordance-centric diffusion policy, we integrate the precise control of imitation learning with the generalization of foundation models.

Related papers:

Hu Site, Takayuki Nagai, Takato Horii. “ ”. IEEE Robotics and Automation Letters, August 2025.

Embodiment Selection and Distance-Aware World Models for Mobile Manipulation

Mon, 01 Jan 2024 00:00:00 +0000

We study world-model-based reinforcement learning for mobile manipulation that jointly addresses embodiment selection (when to move the base vs. use the arm) and motion planning. By introducing reachability rewards and distance-aware rewards, we move beyond similarity-based latent-space rewards toward reliable motor-skill acquisition for separately controlled mobile-manipulation behaviors.

Related papers:

Xiaoxu Feng, Takato Horii. “ ”. IEEE Access, Early Access, 2026.
Xiaoxu Feng, Takato Horii, Takayuki Nagai. “ ”. IEEE Robotics and Automation Letters, February 2025.
Xiaoxu Feng, Takato Horii, Takayuki Nagai. “Predictive Reachability for Embodiment Selection in Mobile Manipulation Behaviors”. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025.

Foundation Models for Humanoid Motion Learning

Mon, 01 Jan 2024 00:00:00 +0000

We apply foundation models such as GPT to humanoid motor control, building unified policies that handle diverse motion tasks including locomotion and whole-body actions. By leveraging the general-purpose reasoning of pretrained models and refining them with action data, we aim at data-efficient, multi-task humanoid control.

Related papers:

Siddharth Padmanabhan, Kazuki Miyazawa, Takato Horii. “ ”. IEEE Access, Vol. 14, 2026.
Siddharth Padmanabhan, Kazuki Miyazawa, Takato Horii, Takayuki Nagai. “ ”. IEEE Access, Vol. 13, 2025.

Multimodal Instruction Foundation Models with Mask Images

Mon, 01 Jan 2024 00:00:00 +0000

We extend vision-language-action (VLA) foundation models with attention-guided mask images, making the correspondence between natural-language and visual instructions explicit so that robots can accurately capture instruction targets and act on them. Through Alpha-channel attention control, we aim at action selection consistent with both language and vision even in cluttered environments.

Related papers:

Shogo Yanagida, Tatsuya Aoki, Tadahiro Taniguchi, Takato Horii. “Mask-guided VLA: Introducing Vision-Language Instructions with Attention-Guided Mask Images” (in Japanese). Annual Conference of the Japanese Society for Artificial Intelligence (JSAI), June 2026.
Shogo Yanagida, Tatsuya Aoki, Tadahiro Taniguchi, Takato Horii. “A Robot Foundation Model Understanding Language and Visual Instructions via Alpha-Channel Attention Control” (in Japanese). 43rd Annual Conference of the Robotics Society of Japan, September 2025.

Whole-Body Teleoperation and Data-Collection Infrastructure for Humanoids

Mon, 01 Jan 2024 00:00:00 +0000

To support foundation-model learning on humanoids and quadruped robots, we develop whole-body teleoperation interfaces and data-collection environments. From affordable exoskeletons for whole-body teleop, to semi-autonomous data collection guided by behavior trees, to evaluation benchmarks for teleop systems, we strengthen the infrastructure that underpins the quantity and quality of action-learning data.

Related papers:

Kazuki Miyazawa, Yuta Kuroda, Takato Horii. “Evaluation Benchmarks for Whole-Body Teleoperation of Humanoids” (in Japanese). 43rd Annual Conference of the Robotics Society of Japan, September 2025.
Kaito Iwamoto, Kazuki Miyazawa, Takato Horii, Hiroshi Ishiguro. “Behavior-Tree-Based Data-Collection Support for Robot Foundation Models” (in Japanese). 43rd Annual Conference of the Robotics Society of Japan, September 2025.
Yuta Kuroda, Kazuki Miyazawa, Takato Horii. “MAITET: Development of an Affordable Modular Multi-Joint Whole-Body Exoskeleton for Humanoids” (in Japanese). Robotics and Mechatronics Conference 2026, June 2026.
Yusuke Yamada, Tatsuya Aoki, Kazuki Miyazawa, Kensuke Iwata, Takato Horii. “A Teleoperation System Supporting Whole-Body Manipulation Learning for Quadruped Robots” (in Japanese). 43rd Annual Conference of the Robotics Society of Japan, September 2025.