Vittorio
Pastore

Robotics Researcher at Virginia Tech.
Passionate about Reinforcement Learning, Music, and Photography.

Publications

A white quadruped robot mid-stride, carrying a robot arm on its back that reaches forward

IEEE RA-L

A Nonlinear MPC Framework for Loco-Manipulation of Quadrupedal Robots with Non-Negligible Manipulator Dynamics

A real-time NMPC framework for quadrupedal loco-manipulation that couples simplified locomotion templates with full-order manipulator dynamics to effectively handle non-negligible arm mass.

Read paper

Three quadruped robots, each standing inside its own glowing amber safety ring on the floor

Under Review

ADMM-Based Distributed MPC with Control Barrier Functions for Safe Multi-Robot Quadrupedal Locomotion

A fully decentralized MPC framework using ADMM and control barrier functions to enable real-time, safe trajectory planning for multi-agent quadrupedal teams with significantly reduced computational overhead.

Read paper

Three quadruped robots following dotted amber paths; two of them trade places on paths that bow around each other

In Progress

Multi-Agent Navigation via Distributed Consensus-Based Optimization and Learned Locomotion

A hierarchical control architecture for multi-agent legged navigation that integrates high-level distributed NMPC using ADMM for scalable collision avoidance with a low-level reinforcement learning policy for robust real-time locomotion.

Coming soon

Three small quadruped robots walking across the cover of a large hardback book

VT Library

Master's Thesis

Thesis research focused on multi agent locomotion via blending of classical and learning-based controls.

Read thesis

Projects

Training a Humanoid on a MacBook

A world-model motion tracker ported from NVIDIA GPUs to MLX: 1,024 simulated Unitree G1s on an M1 Pro, and one policy that tracks 40 motion clips after 79 minutes.

PythonMLXMuJoCoReinforcement Learning

Read the post

A crowd of small G1 humanoid robots moving across the keyboard of an open MacBook whose display shows a training curve

131 Hours of Motion on an RTX 5070

The full 131-hour BONES-SEED motion library, trained on with one RTX 5070: a mid-range consumer GPU with 12 GB of memory, where the same run used to need 40 GB.

PythonJAXMuJoCoReinforcement Learning

Read the post