Recent papers / robotics / 6 essays
Short write-ups of new robotics papers that change how I think about the field. Each one says what the paper does, what it measured, and what I would check before trusting it.
With no real robot data, only simulated demos, MolmoBot doubles the real-world success of a policy trained on 10,000+ hours of real robot data.
sim-to-realsimulationimitation-learningvla
Keep every stream, compress it as you go, and learn what to remember.
robot-memoryvla
Empowerment measured from unlabeled data guides online exploration.
reinforcement-learningexploration
Faster-WAM cuts a world action model's action expert to one layer. I added it to LeRobot behind opt-in flags and timed it on an A100 with random weights.
world-action-modelslerobotkv-cacheflow-matching
ASPIRE lets a coding agent debug robot programs from execution traces and save each verified fix as a text skill. What the loop does, and what the numbers cover.
code-as-policyagentsskill-learningnvidia
Human motion guides its experiments. Every joint stays free.
reinforcement-learningexplorationdexterous-hands