Conference Presentation
Why Robotics Still Isn't Solved - But Could Be Soon | YC Paper Club
Y CombinatorFrancois Chaubard, Marcel Torne, Milan Ganai, Tyler Ga Wei Lum, Niko West, Bill Jiao, Guanming Wang, Shyam Gollakota
- Gary Tan predicts that robotics will be "solved" within the next year, a claim repeated for a decade, while expressing uncertainty about 2026 being the "year of robotics" and asserting it will at least be the "year of demos" if full deployment does not occur by the end of 2026.
- Significant technical risks include the reliance on scaling tele-ops data collection leading to industry failure, sim-to-real models that do not respect physics, action space representations that remain unsolved, and embodied drift from actuator corrosion and battery degradation requiring VLAs to be retrained as mappings become stale.
- Current limitations prevent robots from performing sensory-motor tasks like finding objects by touch due to a lack of epidermis, and precise assembly skills cannot be learned from scratch without first mastering object manipulation in free space.
- Nico projects that robotics application companies will become the "new SaaS," driving economic transformation and building defensible business moats against pure model companies that will not perform necessary operational work.
- The physical world presents unpredictable failure modes that cannot be fully anticipated in the lab, necessitating workarounds and storage layers for physical data to avoid excessive engineering overhead.
- Tyler Lum reports that SimToolReal can execute dexterous tasks requiring multi-fingered hands zero-shot across novel tools and tasks, demonstrating strong recovery behaviors when dropping tools, though it currently cannot simulate objects like water or zippers and struggles with continuous arm twisting.
- Bill highlights that World Action Models (WAM) currently require two GB200s and cost approximately $70,000, making them economically unscalable, but predicts distilling flow matching from 50-100 steps to one or two will yield a 50x speedup without performance loss.
- Optimized infrastructure is expected to allow WAM to run 500 milliseconds per chunk and perform real-time on the edge, enabling the model to learn correlations between pixel-level physics and actions, though WAM still requires real-world fine-tuning for priors that cannot be simulated.
- Nico asserts that successful startups do not require billion-dollar seed funding and that teams must determine data scaling curves through active training rather than upfront estimation, while others note that solving teleop problems generally allows for model training.