Preprints & reports
Research I'm developing and technical reports I've written. The full write-up (figures, equations, tables, and references) is linked from each.
Technical report
Neural Cup Pong: A Playable Action-Conditioned World Model
How small can a genuinely playable world model be, and what breaks first as you shrink it?
A 409K-parameter GRU dynamics model and a 262K-parameter state-grounded decoder run an entire 3D cup-pong game from controller input, with no game engine and no graphics renderer in the loop. The learned model nails control (aim/power errors ~0.002, 99.4% phase accuracy, 100% directional response) but hits a "flight-precision wall": a small structured model cannot localize a ballistic landing to within a cup radius. A hybrid seeds exact flight from the network's own accurate control, restoring engine-parity sinking (6/6) while control and rendering stay fully neural.
Preprint
Diffusion-Trained Backbones for Flow-Matching VLA Policies
Does a diffusion-trained backbone produce better action-token hidden states than a matched autoregressive one when both drive the same flow-matching action expert?
A controlled LIBERO study isolating the pretraining prior with a matched pair: Dream-7B (diffusion) vs. Qwen2.5-7B, its own autoregressive base, driving one shared flow-matching expert. The diffusion-trained backbone improves mean LIBERO success by +5.9 points (82.2 vs. 76.3), concentrated on the long-horizon (+12.0) and object (+11.2) suites where the AR backbone is weakest, and a tie on the near-ceiling suites. Linear probing finds the two equally decodable for low-level action, state, and gripper targets, so the behavioral gain does not reduce to the linear readability of immediate control. Single-seed, simulation-only; a controlled existence proof.
Research proposal
Origami-MoH: Tactile-Augmented Mixture-of-Horizons Flow Policies for Bimanual Dexterous Origami
On a 65-DoF bimanual dexterous platform with rich tactile sensing, does augmenting a flow-matching VLA with a Mixture-of-Horizons reactive head and progressive tactile fusion improve autonomous paper folding over a vision-only policy?
A π0.5 flow policy made reactive (multi-horizon action prediction with disagreement-based replanning) and contact-aware through progressive tactile fusion, escalating from proprioceptive concatenation to a force-aware mixture-of-experts action decoder. Trained as a single vision-conditioned policy over a fixed six-fold sequence across 143 tactile-rich sessions, with a staged Phase 0–4 plan whose go/no-go criteria isolate the value of tactile sensing before the most invasive fusion is built, plus the systems work to fine-tune a 65-DoF policy from a 32-DoF base in OpenPI.