This repo contains tutorial notebooks for the Hyperdimensional Computing (HDC) + graph demo in ODSC West 2026.
The demo shown in our talk follows a fleet of warehouse robots that learn from each other using hypervector representations. When R17 has a "near-miss" (avoided collision) with a forklift in aisle 8, or R04 skids on oil in aisle 9 and finds that slowing down lets it cross the obstacle, those experiences become part of a memory the whole fleet shares, so a new robot R11 already knows to slow down before it enters aisle 9 as it encounters the oil spill for the first time.
The full demo shown on stage is a real running system built with HDC: a fleet simulation with simulated sensors and LiDAR, a policy engine that picks each robot's response to a hazard, a REST API service with a UI that watches the fleet move in real time, and a hypervector memory that is revised episode by episode as the robots learn. The code for that system is complex, and isn't included in this intro tutorial repo.
The notebooks in this repo are designed to introduce the basics of HDC, and the following key aspect of the workflow we built:
HDC proposes candidates. The graph validates whether those paths actually exist.
The code walks through use a simple warehouse graph from the talk (3 robots, 3 aisles, 3 hazards and 2 responses) and just three libraries: LanceDB for storage, lance-csr for Cypher queries, and torchhd for HDC. We don't complicate this example: there are no sensors, no simulation and no UI, so every step fits in a few lines of code. They assume you know graphs and graph databases, and nothing about HDC.
- A warehouse robot graph in LanceDB: store the graph as LanceDB tables and query it with Cypher. This is the exact source of truth.
- The three operators of HDC: hypervectors, and the three operations on them, binding ⊗, bundling ⊕ and permutation ρ, with the math beside each one.
- HDC proposes, the graph validates: encode the whole graph as one hypervector, ask it questions, and check every proposed answer against the graph, including the wrong ones.
You need uv. It installs Python 3.12 for you if it's missing.
On Apple Silicon Macs (macOS 11+) and Linux x86_64 (glibc 2.28+, such as Ubuntu 20.04+),
lance-csr installs from a prebuilt wheel. No Rust or protoc installation is needed.
git clone https://github.com/hyperdimensionalcomputingai/odsc-west-2026.git
cd odsc-west-2026
uv sync --locked --no-build-package lance-csr
uv run jupyter lab src/The install command requires a wheel for lance-csr, so it won't unexpectedly start a source
build. Intel Macs and Windows do not have prebuilt lance-csr wheels; use one of the supported
platforms for this tutorial. For source builds, see the
lance-csr installation guide.
Open 01_warehouse_trial_graph.ipynb first. On Linux, torch installs as a CPU-only build, so
there's no multi-gigabyte CUDA download.
Run the tests (executes every notebook headless) with uv run pytest.