Ranulfo Bezerra
Research

05

Edge Physical AI

Reasoning that cannot run on the robot is not reasoning the robot can use.

The other four pillars are constrained by what a physical agent can actually compute, in time, on a power budget, in the field. This pillar develops compact models grounded in graph representations, distributes inference across the robot-IoT-edge continuum, and validates in simulation before hardware.

Open questions

  1. 01

    What is the smallest model that can reason usefully over a physical knowledge graph?

  2. 02

    How should inference be split across a robot, a sensor node, and an edge server?

  3. 03

    What transfers from simulation, and what never will?

01/Methods and topics

02/Research funding

03/Related projects

04/Representative publications

13

2024

Backhoe Downhill Movement in Varied Earth Structures using Deep Reinforcement Learning

Ranulfo Bezerra, Kazunori Ohno, Shotaro Kojima, Satoshi Tadokoro

RSJ 2024·Domestic conference

2023

Deep reinforcement learning for backhoe downhill movement on variational soil density

Ranulfo Bezerra, Yuta Suzuki, Kazunori Ohno, Shotaro Kojima, Satoshi Tadokoro

RSJ 2023·Domestic conference

2025

Hierarchical Semantic 3D Scene Graphs Generation for Industrial Simulation Environments by Using LVLMs

Nathan Walus, Ranulfo Bezerra, Kazunori Ohno

ROBOMECH 2025·Domestic conference

2025

Multimedia Source Integration Using Specialized LLMs for News Generation

Ahmed Youssef, Ranulfo Bezerra, Tsige Tadesse Alemayoh, Shotaro Kojima, Kazunori Ohno

ROBOMECH 2025·Domestic conference