P9 · the body
DARS / Physical AI
Distributed robots, swarms, embodied evolution and sim-to-real. Sensorimotor ground, at collective scale.
Research · Faber
Nine programmes for evolving intelligence on the BEAM
Faber is the evolutionary layer of the stack: neuroevolution built to run populations across a federated mesh. Its research is organised as an open, signed corpus — every finding, positive or negative, written down and grounded in the literature. These are the programmes that corpus is built around.
The shape
One engine, two couplings
Seven programmes are the evolutionary engine — the near-orthogonal axes of the design space. Two more couple that engine to the world: to language, and to the body. It is the symbol-grounding problem drawn as a map.
P1 – P7 · the engine
The evolutionary engine
Representation, search, meta-adaptation, objectives, scale, encoding and coevolution — the seven axes that hold steady while each is varied in turn, so findings compose.
P8 · language
LLM-augmented Evolution
Language models as variation operators and objective authors. The bridge to language-model cognition.
The programmes
The nine charters
Each programme has a charter: its axis, its open questions, a first experiment, and where the literature already stands. Four engine programmes are characterised (an arc of signed findings reached a synthesis); coevolution is the active front; the rest are chartered and open to take up. No programme is ever closed: "characterised" is a statement about where attention is, not a verdict that the questions are finished. Any programme reopens the moment new evidence, a better instrument, or a sharper question arrives.
P1
Characterised
Capabilities
Representation — what the network can express
Which network capability (tuner depth, memory mechanism, topology) helps which problem. Arc: capability value = task-match x budget x optimizer.
Read the charter
P2
Characterised
Search Strategies
Operators — variation and selection
Which search algorithm (GA, ES, CMA-ES, CoSyNE) suits which landscape, and whether a stronger optimizer overturns the capability findings.
Read the charter
P3
Characterised
Meta-learning
Meta-adaptation — learning to learn
Evolved plasticity and lifetime learning. Signed arc: memory can come from learning rather than storage, and lifetime learning is a shallow interaction window, not deep integration.
Read the charter
P4
Characterised
Objectives
Selection pressure — what fitness rewards
Quality-Diversity and novelty search. Signed arc: novelty illuminates a diverse archive, solves a deceptive maze where objective search and a strong optimizer are trapped, and buys that coverage at no measurable speed cost where the objective already wins.
Read the charter
P5
Open
Scale / Substrate
Distribution of the search — mesh and federation
Island models, distributed populations, and federated evolution across independently owned mesh nodes. Where the substrate becomes the subject.
Read the charter
P6
Open
Encoding / Development
Genotype to phenotype — direct vs generative
Direct versus indirect and developmental encodings (CPPN, HyperNEAT, cellular encoding). The axis that gates scale-up.
Read the charter
P7
Active
Coevolution / Self-play
Interaction — fitness from contest
Coevolution, arms races, open-endedness: the problem moves as agents interact. The active front, climbing a gradual ladder from numbers games (where progress has a known answer) toward embodied predator-prey. First finding: co-fitness is blind to progress, so a fixed benchmark is needed.
Read the charter
P8
Horizon
LLM-augmented Evolution
Augmentation — language in the loop
Language models as intelligent variation operators and objective authors; the bridge between neuroevolution and language-model cognition.
Read the charter
P9
Horizon
DARS / Physical AI
Embodiment at scale — distributed robots
Distributed Autonomous Robotic Systems: swarms, embodied evolution, and sim-to-real, with the mesh as the coordination fabric.
Read the charterWhere it stands
One wall, three explanations — and what the fork found
Two programmes converged on the same open question: a classic non-Markov control benchmark that neither the genetic algorithm nor a self-adaptive evolution strategy solved. That single wall had three competing explanations, one on each of three axes. The fork has now been run, and the answer is not the tidy one we expected:
- P3 Representation — the memory timescale. Still live: a real factor, but not the whole wall.
- P2 Optimizer — search strength. The wall was localised to optimizer-plus-representation together, not to any single axis.
- P4 Deception — ruled out for this wall as a signed negative, then confirmed as a real phenomenon on a purpose-built deceptive maze, where novelty search solves what objective search and a strong optimizer cannot.
One leg became a signed positive elsewhere, one a signed negative here, and the benchmark itself resolved into a joint optimizer-and-representation limit. The negatives were as load-bearing as the positive — which is the point of writing them all down.
Take part
Pick a charter, run an experiment, sign a finding
The corpus is a commons. The harness is built and green; the open questions are small enough to start on in days. An open charter names its first experiment against tools that already run.
1
Read the charter
Each open programme states its axis, its first experiment, and where the literature already stands, so you begin with grounding rather than a blank page.
2
Run against the harness
The scapes, the evolutionary loop, and the plan-feed-insight cycle are in place. A first experiment is a bounded run, not a new framework.
3
Sign the finding
Write the result into the shared corpus — evidence cited, negatives kept. Findings are reported as reproducible replication, never overclaimed as discovery.