Research
Technical checkpoints, algorithm notes, and hardware milestones as we build the adaptive-support platform.
NS-0.1: baseline models for warehouse pick-and-place
Our first public checkpoint on the internal manipulation stack. Honest throughput and repeatability on real cells—not leaderboard scores on synthetic scenes.
Read →February 28, 2026Memory-augmented policies for long-horizon tasks
Long tasks fail when policies have no notion of what happened sixty seconds ago. We add a short-term episodic buffer and a small latent cache for recurring subgoals.
Read →January 15, 2026Tokenization for continuous robot actions
Continuous actions can be discretized with piecewise-linear bins or learned vocabularies. We compare inference latency and tracking error on a 7-DOF arm doing contact-rich insertion.
Read →December 2, 2025From simulation to hardware without domain shift
The gap between sim and floor is often calibration, not graphics. We document a loop: time-align actions, match contact modes, then freeze a small set of gains per station.
Read →October 20, 2025First generalist stack for Nanoshape workcells
High-level map: perception (multi-camera + depth), planning (short-horizon MPC + library moves), and safety monitors in one deployable bundle.
Read →October 13, 2025Gradient Swapping for Effective Agent Reconfiguration?
We study a distributed ensemble where each agent maintains a local hop estimate and occasionally proposes a pairwise goal swap when both sides reduce Manhattan cost.
Read →