current
last updated
curious about
keeping a failure detector calibrated inside a loop that keeps patching the policy it monitors. whether that failure signal generalises across tasks and embodiments or stays stubbornly per-task. and how far you can trust a world model's imagined rollout before it drifts from real.
building
a hierarchical failure detection setup on a cheap so-101 arm, where weak signals escalate to expensive ones only when something looks off. on the side, reproducing a jepa-style world model from scratch and digging into exactly where it breaks instead of chasing the paper's numbers.
reading
sadly just research papers, not getting time to read anything else
training for
delhi half marathon on october 18th. trying to increase my weekly running mileage from 15-20k to 25-30k