Andrea Tagliabue

I am an Applied Scientist working on robotic manipulation at Amazon Robotics, where I develop model-based and learning-based methods for perception, planning, and control in real-world systems. I received my Ph.D. from LIDS at MIT in 2024, focusing on imitation learning for robust, adaptive, agile sensing and control under uncertainty. My research lies at the intersection of learning and control, spanning foundation models for robotics, data-efficient imitation and reinforcement learning, and learning-based model predictive control. Prior to MIT, I was a robotics engineer affiliate at the NASA Jet Propulsion Laboratory and Caltech, contributing on state estimation for the DARPA Subterranean Challenge. I earned my M.S. in Robotics from ETH Zürich, was a visiting researcher at UC Berkeley, and received my B.S. in Automation Engineering from Politecnico di Milano.

[andrea.tagliabue@me.com | LinkedIn | Google Scholar]

Selected Publications

On a 750 mg flapping-wing robot, this controller reaches 197 cm/s and 11.7 m/s² saccades (447% and 255% above prior results), tolerates 160 cm/s wind, and executes 10 flips in 11 s.

[paper | video]

This method enables zero-shot transfer from a policy trained on a single robust-MPC demonstration, substantially reducing data collection and training effort while preserving robustness to unseen perturbations.

[paper | video]

Tube-NeRF reports an 80x gain in demonstration efficiency, 50% lower training time, and only 1.5 ms onboard inference for robust vision-based multirotor tracking.

[paper | video]

NeBula was fielded across wheeled, legged, and aerial robots, and supported CoSTAR’s 2nd-place (Tunnel 2019) and 1st-place (Urban 2020) DARPA SubT finishes.

[paper | video | project webpage]

The adaptive policy is learned in about 1.3 hours and achieves 6.1 cm average position error under winds near 50% of robot weight, including disturbances 36% beyond training.

[paper | video]

On a 0.7 g SoftFly platform, this controller achieves <1.8 cm position RMSE in challenging maneuvers and reduces maximum position error by 60% versus prior work.

[paper | video]

REAL integrates LLM reasoning into real multirotor mission and control loops at 0.1-1.0 Hz, reducing tracking error under controller-parameter mismatch and unmodeled dynamics.

[paper]

LION fuses high-rate IMU and low-rate lidar in a fixed-lag smoother with online extrinsic calibration and observability-aware confidence, enabling robust field deployment in GPS-denied subterranean settings.

[paper | video]

This work demonstrates robust collaborative payload transport by multiple mechanically coupled MAVs without explicit inter-agent communication, validated through extensive simulation and flight experiments.

[paper | video | teaser video]

CGD combines diffusion policies with online surrogate optimization to generate collision-free, dynamically feasible trajectories and improves feasibility/performance under unseen constraint changes.

[paper]