The Frontier of Robotics Research | Bessemer Venture Partners | Robotics Day

Why It Matters

Robotics progress now hinges on mastering high‑quality, scalable data and unified stacks, making them the decisive competitive edge for investors and enterprises alike.

Key Takeaways

• Data quality and quantity remain the biggest bottleneck in robotics.
• Full‑stack approaches prioritize infrastructure over isolated foundation models.
• Companies balance in‑house hardware development with outsourced standardized components.
• Reinforcement learning struggles in physical AI; reward design is an art.
• Scaling robotics models requires new architectures beyond traditional vision transformers.

Summary

The Bessemer Robotics Day panel brought together three AI‑research‑origin founders—Jason of Dyna, Philip of Exos Dof, and Armin of Perceptron—to map the current frontier of embodied AI. Moderated by partner Janelle, the discussion centered on how data, hardware, and learning algorithms intersect across the robotic stack.

All three speakers emphasized that data is the limiting factor. Philip described a "data pyramid" where high‑quality, robot‑specific data sits at the top but is scarce, while abundant, lower‑fidelity data lives at the base. Dyna’s strategy is to invest heavily in end‑to‑end infrastructure that captures, filters, and validates data before feeding foundation models, arguing that without reliable data no model can succeed. Perceptron’s Armin highlighted that traditional reinforcement learning (RL) often fails in physical settings, making reward‑shaping an art and prompting new formulations that synthesize data rather than directly train policies.

Concrete examples illustrated these points: Jason allocated a hypothetical $100 budget primarily to data capture and pipeline development, noting that off‑the‑shelf components like cameras are outsourced while critical robot hardware and models stay in‑house. Philip noted that egocentric datasets such as Ego4D have grown more useful as humanoid platforms mature, illustrating how data relevance evolves. Armin shared that his team is preparing a paper on novel RL formulations that use verifiable rewards for synthetic data annotation, underscoring the shift toward hybrid approaches.

The panel’s consensus signals a strategic pivot for robotics startups: prioritize integrated data pipelines, build bespoke hardware where performance gaps are large, and develop new model architectures tailored to multimodal, video‑rich inputs. Investors and enterprises should watch for firms that can close the loop between real‑world deployment and continuous data improvement, as these capabilities will dictate the speed of commercial robot adoption.

Original Description

Three founders at the cutting edge of embodied AI join Bessemer partner Janelle Teng Wade for a candid discussion on what's actually working in robotics research — and what isn't.

Janelle is joined by

— Jason Ma, Co-Founder of Dyna Robotics (full-stack manipulation);

— Philipp Wu, Co-Founder of a stealth startup (data infrastructure and teleoperation);

— Armen Aghajanyan, Co-Founder and CEO of Perceptron (vision-language models for physical AI).

Together, they cover:

— The data bottleneck: why robotics is still orders of magnitude behind language models, and how each company thinks about the tradeoff between teleoperation data, egocentric video, simulation, and internet-scale pre-training

— Build vs. buy decisions: what to own in-house versus outsource when you're building a full-stack robotics company, and why the gap between 80% and 99% reliability changes everything

— Reinforcement learning in robotics: why traditional RL is harder to apply than in LLMs, how to build effective reward signals, and why human demonstrations are still the primary bottleneck before RL can unlock superhuman performance

— Hot takes: overhyped vs. underhyped: pixel-level world model reconstruction, physics understanding, and data attribution as an overlooked research frontier

— Who's doing impressive work: Physical Intelligence, Chinese open-source labs, Unitree, and the academic community's continued role in hardware innovation

This is a must-watch for robotics founders, researchers, and engineers building at the intersection of AI and the physical world.

Get more robotics research when you subscribe: bvp.com/subscribe