The frontier of robotics research | Bessemer Venture Partners | Robotics Day
Janelle, a partner at Bessemer Venture Partners, moderates a Robotics Day panel with founders Jason of Dyna, Philip of Exos Dof, and Armin of Perceptron on robotics data, model training, reinforcement learning, and hardware.
In this talk
- 1:26
Robotics data and embodiment
Philip says language models benefit from abundant training material, while robotics still has a much smaller pool. He describes a tradeoff: data that closely matches a target robot is valuable but hard to gather, while easier-to-collect data is more plentiful but less specific to that robot. Jason adds that data’s usefulness can change as robot hardware improves, pointing to Ego4D’s 4,000 hours of footage released in 2020.
- 7:11
Building the robotics stack
Jason says Dyna expanded toward hardware and infrastructure after finding that business workflows need more than two fixed arms. He prioritizes a pipeline that captures, filters, and validates data through model training and deployment. Dyna buys standardized components such as cameras, while developing its own end effectors, robots, and models where existing products do not meet its needs.
- 13:10
Training models and using reinforcement learning
Armin says conventional reinforcement learning is difficult to apply directly to physical robot policies because rewards are hard to verify. He describes using RL to improve synthetic data annotation and using detailed reasoning traces to make policy training more effective. Panelists say RL already supports locomotion and sim-to-real control, while manipulation still needs better human demonstrations and reward signals; visual language models can offer a starting point for measuring task progress.
- 19:12
Research gaps and influential work
Armin questions pixel-level reconstruction as the main route to world models and says robotics needs stronger end-to-end training methods. Panelists also highlight physical system characterization and data attribution as underexplored areas: teams often cannot trace a robot’s behavior to specific training examples. They point to Physical Intelligence, Chinese research labs, Unitree, and academic work such as the UMI data-capture research as useful sources of ideas.
- 25:54
Data sources and video model backbones
Panelists see video-pretrained models as a promising direction because their real-world visual training may better match what robots encounter. They distinguish internet video and open datasets from teleoperation data, which can include robot actions, proprioception, tactile input, audio, and multiple camera views. Armin says open datasets may suit mid-training, while robot-specific data can be more useful for final policy tuning.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.