research

Papers.

Mostly one obsession: getting robots to work with less — less labelling, fewer sensors than you'd think you need, and honest numbers about what actually helps.

ScaffDiff

SMC 2026 · accepted · RA-L · in revision

A LiDAR sweep sees maybe a twentieth of the scene. ScaffDiff completes the rest — walls, ground, the parked car — in one denoising step, about 209 ms per frame where prior diffusion methods take up to 30 seconds.

drag to rotate

FIG. 1 — sparse sweep in, dense scene out, one step. 65–138× faster than diffusion baselines on the same GPU.

Diffusion models usually buy quality with dozens of denoising iterations. The surprise in this work: for scene completion, the scaffold — where you put the initial coordinates — matters far more than how long you denoise. Get the scaffold right and a single step suffices; the encoder barely matters, the sensor modality doesn't either. Two papers: the SMC one (accepted) builds the evaluation substrate, the RA-L one makes the scaffold-dominance case.

HapticVLA

IROS 2026 · accepted

Tactile sensors are great in the lab and annoying on real robots. HapticVLA uses touch only during training — then distills it into a student that handles an egg with no tactile sensor on board, predicting the touch signal from what it sees.

switch policies — same egg, same gripper, no sensor either way

FIG. 2 — the plain policy squeezes blind past the crack line. the distilled student predicts the contact force (dashed) and settles in the safe band — sensor-free.

Two ingredients: a teacher trained with a safety-aware tactile reward that penalizes over-force and slip, and a distillation step that teaches a standard vision-language-action model to predict the tactile signal it no longer receives. Evaluated on real fragile-object tasks — eggs included. In review; numbers and write-up when the process allows.

Phantom

ICRA 2027 · in progress

You know the glass is about to slip a beat before it does. Phantom is a world model that gives robots that beat: it announces contact and slip before they happen, and slows the arm down when its own imagination gets uncertain.

FIG. 3 — hollow marks: what the model says is coming. filled marks: what the world then does. the gap is lead time — and high uncertainty visibly slows the motion.

The bet: a robot that anticipates contact doesn't need to sense it everywhere — rich touch can live in training and become prediction at deployment, with the model's own uncertainty acting as the speed governor. Early days; details when there's something worth defending.