Data for Continual Improvement
- Deployment, interaction, and failure data
- Human feedback and correction data
- Data curation and the role of pretraining
Overview
A robot policy at deployment is not the finish line; it is a new starting point. In this workshop, we are interested in improvement in deployment of robot policies: how robots get better over time through interaction, feedback, and adaptation. This workshop aims to bring together researchers working across the full lifecycle of robot improvement, from how data is collected and curated, to how policies are updated, adapted, and evaluated in the real world.
Central questions include what data we collect, how we use it, what pretraining enables, including increasingly powerful vision-language-action (VLA) models, what infrastructure supports rapid and reliable iteration, and how progress should be evaluated.
By bringing together researchers with different expertise across this stack, we hope to create a space to discuss the core scientific and systems questions behind this vision, and to help shape a more unified perspective on how robots improve over time.
Topics of Interest
Algorithms for Improvement
- RL finetuning, reward modeling, and preference alignment
- Continual learning and test-time adaptation
- Improving and adapting foundation and VLA models
Systems for Real-World Iteration
- Infrastructure for real-world policy iteration
- Human feedback collection at scale
- Fleet learning and multi-robot systems
Benchmarking & Evaluation
- Metrics for adaptation speed and robustness
- Safety and forgetting during learning
- Long-horizon evaluation in deployment
Speakers
Speaker announcements coming soon.
To be announced
Industry
To be announced
Industry
To be announced
Academia
To be announced
Academia
To be announced
Academia
To be announced
Academia
To be announced
Academia
Call for Papers
Topics
We welcome submissions on, but not limited to:
- Data collection, curation, and replay for iterative policy improvement
- RL finetuning, reward modeling, and preference alignment for robot policies
- Human-in-the-loop learning and scalable human feedback systems
- Continual learning and mitigating catastrophic forgetting
- Test-time adaptation and in-context learning for robots
- Systems and infrastructure for real-world policy iteration
- Fleet learning and multi-robot experience sharing
- Benchmarks and evaluation metrics for continual improvement
- Safety-aware improvement and exploration
- Improving and adapting foundation and VLA models
Important Dates
Details to be announced.
| Milestone | Date |
|---|---|
| Submission Deadline | TBD |
| Acceptance Notification | TBD |
| Camera Ready | TBD |
| Workshop | TBD |
Program
Morning
8:45 – 9:00 AM
Welcome
Opening Remarks
9:00 – 9:30 AM
Invited Talk 1
Industry
TBA
9:30 – 10:00 AM
Invited Talk 2
TBA
10:00 – 10:30 AM
Invited Talk 3
TBA
10:30 – 11:15 AM
Break & Posters
Coffee Break + Poster Session
11:15 – 11:45 AM
Invited Talk 4
TBA
11:45 AM – 12:15 PM
Poster Spotlights
Spotlight Talks 1, 2, 3
12:15 – 1:15 PM
Lunch
Lunch Break
Afternoon
1:15 – 1:45 PM
Invited Talk 5
TBA
1:45 – 2:15 PM
Invited Talk 6
Industry
TBA
2:15 – 2:45 PM
Invited Talk 7
TBA
2:45 – 3:30 PM
Break & Posters
Coffee Break + Poster Session
3:30 – 4:00 PM
Poster Spotlights
Spotlight Talks 4, 5, 6
4:00 – 4:45 PM
Panel Discussion
TBA
4:45 – 5:00 PM
Closing
Closing Remarks & Awards
Organizers
Ph.D. Student, CMU
Ph.D. Student, Stanford
Ph.D. Candidate, UC San Diego
Ph.D. Student, UT Austin
Postdoc, UC Berkeley
Asst. Professor, CMU