Blog
Thoughts on robotics, AI, research, and technology.
Robot Policies Need the Right Memory, Not an 8K Context Window
RoboTTT, WAM-TTT, and RoboSSM show that long context matters only when history becomes usable state—through task-relevant information, aligned meta-training, and memory matched to its timescale.
How We Won IEEE ICRA Robotics Challenges Two Years in a Row
Lessons from winning ICRA navigation and manipulation challenges: start from the problem, decompose it from first principles, and test the assumptions that matter.
OpenGEN-1: Reconstructing GEN-1 from First Principles
A first-principles reconstruction of GEN-1 as a physical foundation model trained on streams of interaction, with a proposed architecture, objective, data scale, and real-time inference system.
VLA Is Not Dead—At Least Not Yet
DreamZero and WAM split robot policy learning into future-video prediction and inverse dynamics—opening a powerful path beyond VLA, but at a steep deployment cost.
Pi 0.6: Supervised Learning in RL's Clothing
How RECAP reveals that modern RL post-training is fundamentally supervised learning guided by value-based preference signals and environment interaction.
CFGRL: The Theory Behind pi0.6*
How CFGRL reframes RL policy improvement as optimality-conditioned inference, then implements it via classifier-free guidance at sampling time.