Letian Fu

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2026-118

May 12, 2026

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2026/EECS-2026-118.pdf

Modern vision and language models are powerful because they are not trained from scratch for every task. They are pre-trained on large, diverse datasets, adapted through post-training, and increasingly deployed as agents that use tools, feedback, and test-time computation. Robots, however, do not yet benefit from the same scaling paradigm. Robot data is expensive to collect, tied to specific hardware and environments, and difficult to obtain at the scale required for general pre-training. As a result, many robot learning systems remain datahungry, task-specific, and brittle under changes in objects, scenes, embodiments, and task instructions. This dissertation explores how principles from foundation model training and deployment can be adapted to robot learning, with the goal of building robot systems that are more efficient, scalable, and generalizable.

In this dissertation, I study how to make robot learning both efficient and scalable by approaching the problem along three complementary axes: (1) trajectory-level sensorimotor pre-training to improve post-training efficiency and enable few-shot generalization, (2) synthetic robot trajectory data generation techniques that do not rely on human teleoperation of robots, and (3) coding agents for robotic control, strengthened through structured agent harnesses and reinforcement learning with verifiable rewards.

Together, these methods demonstrate how principles from large-scale foundation model training and deployment can be adapted to build general, data-efficient, and scalable robot foundation models.

Advisors: Ken Goldberg

---

BibTeX citation:

@phdthesis{Fu:EECS-2026-118,
    Author= {Fu, Letian},
    Title= {Towards Efficient and Scalable Robot Learning: Trajectory Pre-training, Synthetic Data, and Coding Agents},
    School= {EECS Department, University of California, Berkeley},
    Year= {2026},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2026/EECS-2026-118.html},
    Number= {UCB/EECS-2026-118},
    Abstract= {Modern vision and language models are powerful because they are not trained from scratch for every task. They are pre-trained on large, diverse datasets, adapted through post-training, and increasingly deployed as agents that use tools, feedback, and test-time computation. Robots, however, do not yet benefit from the same scaling paradigm. Robot data is expensive to collect, tied to specific hardware and environments, and difficult to obtain at the scale required for general pre-training. As a result, many robot learning systems remain datahungry, task-specific, and brittle under changes in objects, scenes, embodiments, and task instructions. This dissertation explores how principles from foundation model training and deployment can be adapted to robot learning, with the goal of building robot systems that are more efficient, scalable, and generalizable.

In this dissertation, I study how to make robot learning both efficient and scalable by approaching the problem along three complementary axes: (1) trajectory-level sensorimotor pre-training to improve post-training efficiency and enable few-shot generalization, (2) synthetic robot trajectory data generation techniques that do not rely on human teleoperation of robots, and (3) coding agents for robotic control, strengthened through structured agent harnesses and reinforcement learning with verifiable rewards.

Together, these methods demonstrate how principles from large-scale foundation model training and deployment can be adapted to build general, data-efficient, and scalable robot foundation models.},
}

EndNote citation:

%0 Thesis
%A Fu, Letian 
%T Towards Efficient and Scalable Robot Learning: Trajectory Pre-training, Synthetic Data, and Coding Agents
%I EECS Department, University of California, Berkeley
%D 2026
%8 May 12
%@ UCB/EECS-2026-118
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2026/EECS-2026-118.html
%F Fu:EECS-2026-118