Tai Inui 

We proposed SoftBit, a teleoperation interface that enhances user awareness through real-time soft robot finger shape visualization in extended reality (XR). The system combines proprioceptive sensing with an XR headset (Meta Quest2) to provide users with intuitive visual feedback about finger deformations during manipulation tasks. SoftBiT’s key innovation is a real-time sim-to-real pipeline that estimates and visualizes soft finger shapes, helping users better understand robot-object interactions even when direct visual feedback is limited.

We introduced TriForce Band, a wrist-worn force-myography interface that uses a dense 4×4 triaxial tactile array to capture rich wrist musculature in a compact, wearable form factor. Paired with an MLP-Mixer + SE model, TriForce Band achieves 93.0% accuracy on 10 gesture classes and shows that adding shear components boosts performance by 2.9 pp over normal-force-only input. 

We introduced SoftNash, an entropy-regularized Nash game formulation for non-fighting virtual fixtures in teleoperation. By inflating the robot’s effort weight with a single softness parameter τ, SoftNash smoothly dials assistance from classic high-gain fixtures to agency-preserving, pass-through behavior while maintaining accuracy and reducing conflict and workload in a 6-DoF haptic study. 

We introduced MicCheck, a plug-and-play acoustic contact sensor that repurposes an off-the-shelf Bluetooth pin microphone for low-cost, easily integrated contact sensing in robot grippers. Despite its simplicity, MicCheck achieves 92.9% accuracy on a 10-class material benchmark and boosts imitation-learned manipulation success (e.g., picking–and–pouring from 0.40→0.80), enabling contact-rich tasks like unplugging and sound-based sorting. 

We introduced Geodiffussr, a flow matching-based pipeline that generates realistic, text prompt-guided texture map while strictly respecting an input Digital Elevation Map (DEM). We fuse geometry into the UNet at coarse-to-fine resolutions with Multi-scale Content Aggregation (MCA), allowing the model to respect global landforms and local ridgelines alike.

We present an unsupervised slide quality assessment pipeline that combines seven expert inspired visual design metrics (whitespace, colorfulness, edge density, brightness contrast, text density, color harmony, layout balance) with CLIP-ViT embeddings, using Isolation Forest–based anomaly scoring to evaluted presentation slides. Our results show that augmenting low‐level design cues with multimodal embeddings closely approximates audience perceptions of slide quality, enabling scalable, objective feedback in real time.

Kuma Lab and Google Developer Group on Campus collaborated to host a hands-on workshop on AI. We introduced the basic concepts of AI, through its applications in 3 fields: Computer Vision, Robot Learning, and 3D Reconstruction. 

Under Google Developer Group on Campus, we hosted a hands-on workshop on game development with Unity. After introducing the basic concepts of Unity, we recreated Flappy Bird with the audience. 

The core members of Kuma Lab hosted a workshop on "Robotic Art", where we explored how the two fields that seem to be far apart: Robotics and Art, can go together to form an interesting topic of research.

Under Kuma Lab, I hosted a virtual session on introductory reinforcement learning. This covered basic concepts like Q-learning, Policy Gradient, and Actor-Critic.

I helped demonstrate our lab's caregiving humanoid robot at the Osaka World Expo 2025, which had a total visitor count of more than 25,000,000.

Student ambassador for 技育祭2024 (Geek SAI 2024), the biggest regional tech conference with huge guests like the creators of Ruby and 2Chan.