For each piece the planner extracts vertices where three mutually orthogonal edges meet, and an action translates the moving piece linearly until a chosen pair of corners coincides — a discrete, geometry-derived action space rather than a sampled one. Feasibility is decided by interpolating along the translation and collision-checking against every other active piece, plus a bottom-support test and an arm budget that models one gripper, two grippers, or unlimited support.

From each state the branch tree is pruned at every level, expanded for an n-step lookahead, and each surviving branch finished with a truncated greedy rollout; the move with the best cumulative score is committed. Accepting short-term sub-optimal moves this way is what rescues the goal-distance heuristic from the interlocking structure that traps a purely greedy planner.

The sequence is searched backwards from the interlocked goal and then reversed, because disassembly is the easier direction: every removal moves away from a fully constrained pose, while assembly must find the one entry order that does not deadlock. Plans were generated for one-, two-, and unlimited-arm budgets, with more arms reducing the number of required actions.

The pick-and-place sequence was validated in ABB RobotStudio with RAPID motion programming and transferred to the physical ABB IRB 14000 YuMi. A MuJoCo scene converted from the real robot's URDF (32 bodies, 60 DoF, 16 actuators) carries the six pieces as free bodies for headless rollout. Current work brings gripper access, arm collision and reach into the scoring itself, so a move no arm can perform is rejected at the stage that proposed it.

Presented Dynamic Programming in Cooperative Autonomous Assembly, independent research with Max Gao, arguing that manually sequenced or heuristic assembly plans do not scale to diverse assemblies and that interlocking connections need a formulation reasoning directly about geometry and support. A parallel study asks whether a language model, given solved sub-puzzle traces in context, can propose feasible moves without search — write-up in progress.

A naive implementation computes the full N×N score matrix, writes it out to HBM, reloads it to apply softmax, writes it back, and reloads it once more for the second matmul. On-chip SRAM runs at roughly 19 TB/s against HBM's 1.5 TB/s, so the GPU spends its time waiting on memory rather than computing.

Llama's decoder block adds grouped multi-query attention and a KV cache around RMSNorm, RoPE and SwiGLU. This project swaps only the attention sub-block for a @triton.jit kernel that streams over cached K/V one token at a time, leaving everything else untouched, and benchmarks it against a naive baseline on a 15M-parameter Llama2-architecture model trained on TinyStories. Tested on ASU's Sol cluster (NVIDIA A100) and locally on an RTX 4060.

Five findings across 50 episodes, of which only two justify deleting data — the discipline being to remove data only when the fix is unambiguous and bias-free, and otherwise annotate and defer. The headline measurement: video lags state by 5 frames (100 ms) in every episode.

A three-level schema (episode / segment / frame) with a weak labeler proposing around 10 segments per episode with no human input, so annotators correct proposals instead of drawing boundaries from scratch. Inter-annotator agreement measured at tIoU 0.589.

A config-driven filter registry implementing the audit findings one-for-one, producing a valid LeRobotDataset at 79.7% frame retention, with 16 tests covering alignment and synthetic corruption.