phylo 0.0.0

An open stack to build, RL-post-train, and deploy frontier robot policies.

phylo

An open stack to build, RL-post-train, and deploy frontier robot policies — from pretraining to the real world.

phylo is the unified framework behind the WoRV flagship robot-foundation-model effort. It does not reimplement training or inference; it borrows mature components (HuggingFace Trainer, vLLM, LeRobot data, the vla-eval benchmark harness, starVLA model code) and unifies them behind one light surface.

Keystone: one serving stack is used identically for RL rollout, evaluation, and deployment — parity-tested, so what you train against is what you ship.

Architecture

A dual-system policy (PonderPounce lineage):

• cortex — System-2 Foundation Cortex: a reusable, shared "brain" that reads context (human demonstration, memory, rules, current observation) and emits action-relevant cognition tokens.
• reflex — System-1 execution layer: turns cognition tokens into a specific embodiment's actions. Small, swappable per robot. Adapted via SFT + Reflex RL (RLT-style actor-critic on the frozen cortex's cognition tokens).
• platform — shared infra: the parity serving stack (rollout = eval = deploy), data/checkpoint registries, vla-eval integration.

Training recipe (stages)

[0] action-prior warm-up → [1] cortex pretrain (web-VLM co-training + cross-embodiment robot data) → [2] reflex SFT → [3] reflex RL (+ real-world online)

VLM-handling is a config dimension, not a fixed default — independent tune_vision / tune_llm / tune_projector flags + a Knowledge-Insulation (stop-gradient) option + per-stage freeze transitions; the default follows the imported model's published recipe.

Status

Early scaffold. The Reflex RL prototype (RLT on sim) is validated end-to-end on vla-eval (SimplerEnv × Qwen3-GR00T) — the server lives at vla-evaluation-harness/src/vla_eval/model_servers/starvla_rlt.py and will be folded into reflex/.

Borrows (heavy internals)

HuggingFace Trainer/Transformers · vLLM · LeRobot dataset format · vla-eval (benchmark harness + serving protocol) · starVLA (QwenGR00T model code).