Chelsea Finn
Home > People > Chelsea Finn
Profile
| Field | Details |
|---|---|
| Current Position | Assistant Professor, Stanford University |
| Lab | IRIS Lab (Intelligence through Robotic Interaction at Scale) |
| Company | Physical Intelligence Co-founder |
| PhD | UC Berkeley (Advisors: Sergey Levine, Pieter Abbeel) |
| Undergraduate | MIT (EECS) |
Key Contributions
- Meta-Learning Pioneer: Introduced a new paradigm for few-shot learning with MAML
- ACT & ALOHA: Democratized robot learning with low-cost bimanual manipulation system and learning algorithm
- Physical Intelligence Founding: Co-founded a general-purpose robot AI company
- Robot Learning from Humans: Key figure in human demonstration-based robot learning research
Research Timeline
PhD & Early Career (2013-2018)
UC Berkeley - Advised by Sergey Levine, Pieter Abbeel
| Year | Work | Impact |
|---|---|---|
| 2016 | End-to-End Training of Deep Visuomotor Policies | Image-to-action end-to-end learning |
| 2017 | MAML (Model-Agnostic Meta-Learning) | Few-shot learning breakthrough, 10,000+ citations |
| 2017 | One-Shot Visual Imitation Learning | Learning to imitate from a single demonstration |
| 2018 | Meta-Learning for Robot Learning | Applying meta-learning to robotics |
Stanford Professor (2019-present)
IRIS Lab Founding
| Year | Work | Impact |
|---|---|---|
| 2019 | IRIS Lab Founded | Large-scale robot-human interaction research |
| 2021 | Offline RL Research | Robot learning from offline data |
| 2023 | ACT (Action Chunking with Transformers) | Precise manipulation from 10-minute demonstrations |
| 2023 | ALOHA | $20K low-cost bimanual system |
| 2023 | Mobile ALOHA | Mobile bimanual robot |
Physical Intelligence (2024-present)
Co-founded Physical Intelligence, participated in pi0 development
Major Publications
Meta-Learning
- MAML (ICML 2017) - Most influential paper
- Probabilistic MAML (NeurIPS 2018)
- Meta-Learning without Memorization (NeurIPS 2020)
Robot Learning
- ACT (RSS 2023) - Published with ALOHA
- Learning from Play (CoRL 2019)
- RoboNet (CoRL 2019)
Imitation Learning
- One-Shot Imitation Learning (NIPS 2017)
- Learning to Learn with Compound HD Models (ICML 2019)
Key Ideas
MAML (2017)
Core: Learning initial parameters that can quickly adapt to new tasks with only a few gradient steps
theta* = theta - alpha*grad_theta(L(f_theta')) where theta' = theta - beta*grad_theta(L(f_theta))Impact:
- Standard methodology for few-shot learning
- Applied across robotics, NLP, and computer vision
- 10,000+ citations
ACT & ALOHA (2023)
- Action Chunking: Predicting action sequences instead of single actions
- Low-Cost Hardware: $20K system reproducible in research labs
- LeRobot Default Model: Impact on open-source ecosystem
Philosophy & Direction
Research Philosophy
“For robots to generalize in the real world, they need large-scale diverse experiences”
Research Direction Evolution
- 2013-2017: Deep RL, end-to-end learning
- 2017-2020: Meta-learning, few-shot adaptation
- 2020-2023: Imitation learning, offline RL
- 2023-present: Large-scale robot learning, foundation models
Awards & Recognition
- MIT Technology Review 35 Under 35 (2019)
- NSF CAREER Award
- Sloan Research Fellowship
- ONR Young Investigator Award
- ACM Doctoral Dissertation Award Honorable Mention