RoboBoom
Foxglove· November 13, 2025· 20:17

Actuate 2025: Foundation Reward Models for Robot Learning with Jason Ma

Jason Ma, co-founder of Dyna Robotics, presents at Actuate 2025 on building reliable generalist robot policies, including with Foxglove.

In this talk

  1. 0:22

    Building generalist robot policies

    Ma says Dyna Robotics combines in-house research with commercial deployments so real customer work can guide research and produce production-quality data. The company trains general-purpose manipulation models from teleoperation data, aiming to address the gap between flashy demonstrations and dependable real-world performance.

  2. 3:42

    Demonstrating a range of skills

    A robot brought from Dyna’s Redwood City office folds T-shirts at SFJAZZ without on-site data collection, despite encountering the venue and audience for the first time. Ma also shows tool use, cup stacking, multi-step meal preparation, and candy picking based on changing spoken requests. He says each demonstrated task used fewer than 10 hours of data and shared the same model architecture.

  3. 10:05

    Using reward models to improve reliability

    Ma describes Dyna’s napkin-folding result: more than 850 napkins folded in a continuous 24-hour run at a 99.4% success rate, repeated about four times in one week. He says a standard training approach reached roughly 80%, while a reward model that scores task progress helped identify good and bad trajectories so the robot could learn from its own runs. The model also recovered from errors such as grasping a stack of napkins instead of one.

  4. 16:39

    Open challenges and deployment

    Ma shows the napkin-folding model operating at a restaurant in Los Angeles. He identifies three ongoing challenges: comparing very high-performing models without hundreds of live trials, making useful use of lower-quality human video, and enabling improvement from a small number of trajectories collected at customer sites.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.

More talks