Projects

The session treats artificial life as a form of experimental philosophy. It invites experimental, theoretical, and position papers that use synthetic systems to make assumptions explicit and examine questions about life, agency, emergence, and simulation. Ben Gaskin and Simon McGregor organise it for ALIFE 2026 in Waterloo.

The workshop examines how learning arises and how evolution moves from specific, hard-wired responses to more general and flexible adaptation. It centres on three questions: how to identify learning, which mechanisms enable it, and which ecological conditions support its emergence. Jason Yoder, Anselmo Pontes, Austin Ferguson, and Ben Gaskin organise it for ALIFE 2026.

The ALIFE 2026 paper asks whether a soft body of identical contractile cells can produce edge-sensitive behaviour without assigned sensor or motor roles. A selected cohort of specimens walks to the edge of a plateau, tells a cliff from a goal, and reverses — strongly on the terrain family it was evolved on, only partially on new distances and in other bodies. One specimen, S4, reverses at the cliff with no terrain route into its controller at all. The companion page carries the figures and animations; the code and specimen records are public.

The atlas compares 200 prepared source-code corpora using LZMA normalised compression distance. Each label contributes four repository-disjoint 64 KiB objects; five cumulative distance matrices are embedded in three dimensions and aligned to common anchors. The map shows transfer of byte-level compression regularities, not semantic similarity between languages. The page reports projection fidelity and warns against reading screen-space clusters as facts.

The simulations recreate selected Braitenberg vehicles. Each vehicle couples light sensors directly to motors. Changes in wiring sign and laterality produce movements that observers may describe as fear, aggression, love, or preference. The project makes the explanatory point visible: complicated-looking behaviour can arise from very simple mechanisms.

The project reconstructs Kolmogorov and Uspensky’s 1958 model of computation. The machine stores information in a finite typed complex and computes by exact local replacement around an active node. A worked tree program shows how the abstract rules operate step by step. The implementation follows the original formalism while making the changing memory structure visible.

A small library of translations, gathered around artificial life and the older ideas it descends from. It holds historical work in science and philosophy — Bauer, Kolmogorov, Kraus, Lapicque, Palágyi, Goethe — presented in a reading interface built to look like a code editor.

The library is a JAX-native reimplementation of Evolution Gym, the soft-body robot benchmark of Bhatia, Jackson, Tian, Xu, and Matusik (NeurIPS 2021). It rebuilds the runtime around object-separated simulation: a robot and its terrain compile as separate objects rather than one composited grid, and purely fixed terrain contributes collider and render geometry without entering the dynamic state. The functional design composes with jit, vmap, and lax.scan, so evolutionary runs can evaluate many morphologies at once. It is a public-alpha research library; parity with the original simulator is checked against reference traces under a pinned JAX version, and slopes and freeform terrain remain unsupported.

The Julia platform runs collectives of simple neuron-like nodes in closed sensorimotor loops and measures what they do. There is no central controller and no hand-wired control: the question is whether competent behaviour needs one. Its canonical node is an independent reimplementation of the homeostatic spiking reservoir of Falandays, Yoshimi, Warren, and Spivey (2024), with tracking, Pong, and wall-following as the benchmark tasks. Nodes, tasks, bodies, and measures resolve from registries at run time, and each run is recorded with its resolved plan, seeds, checksums, and tables. Validation covers the declared reference trajectories only, and establishes neither general competence nor biological fidelity.

The repository provides a pure-JAX implementation of Plank and colleagues’ neuromorphic CartPole benchmark. It includes all four difficulty modes, spike encoders, fitness scoring, batched population evaluation, and rendering. Its functional design composes with jit, vmap, and lax.scan, allowing evolutionary experiments to evaluate many controllers and trials in parallel. Reported speedups apply to the specified hardware and benchmark conditions.