baby-model

Research framework for the Baby AD/DA asymmetry hypothesis.

The working hypothesis is:

A model that grows perception first, delays action decoding, and uses prediction improvement as intrinsic reward should form more stable concepts and transfer better than an end-to-end agent trained from the first step.

This repository starts with a small, dependency-free RL smoke environment so the research loop can run on any fleet node. The framework is intentionally small: it gives us a verified loop, run artifacts, tmux entrypoints, and a path to replace the toy environment with MiniGrid, BabyAI, Habitat, or robot data.

Current v0

• A_end_to_end: raw-observation Q-learning baseline.
• B_encoder_first: coarse perceptual representation with a decoder delay.
• C_baby_surprise: coarse representation, decoder delay, and raw intrinsic surprise reward.
• D_baby_progress: coarse representation, decoder delay, and prediction improvement reward.

The v0 environment is not a scientific claim. It is a harness to make the research pipeline testable before we spend GPU time.

Quick Start

./scripts/verify.sh
python3 -m baby_model.cli run --config configs/experiments/v0-smoke.json --output-dir runs
./scripts/launch_tmux_local.sh
./scripts/launch_tmux_sweep.sh

Optional MiniGrid/BabyAI probe:

./scripts/verify_minigrid.sh

This requires the optional minigrid dependency; setup details are in docs/experiments/minigrid-protocol.md.

Optional PyTorch DQN smoke:

MINIGRID_TORCH_CONFIG=configs/experiments/minigrid-torch-unlock-smoke.json \
./scripts/verify_minigrid.sh

This additionally requires optional torch; setup details are in docs/experiments/minigrid-torch-lane.md.

Fleet Loop

Start read-only:

./scripts/fleet_inventory.sh

After the repository is pushed and checked out on worker nodes, use the commands in docs/fleet/worker-plan.md to run local loops under tmux on each host.

Tracking

• Research hypothesis: docs/research/hypothesis.md
• Source notes: docs/research/sources.md
• Experiment protocol: docs/experiments/v0-protocol.md
• v0.2 sweep result: docs/experiments/v02-sweep.md
• v0.3 sweep result: docs/experiments/v03-sweep.md
• MiniGrid/BabyAI migration: docs/experiments/minigrid-protocol.md
• BabyAI Unlock hard task: docs/experiments/minigrid-babyai-unlock.md
• MiniGrid curriculum to BabyAI Unlock: docs/experiments/minigrid-curriculum-unlock.md
• MiniGrid linear function approximation: docs/experiments/minigrid-linear-unlock.md
• MiniGrid linear multi-seed sweep: docs/experiments/minigrid-linear-sweep.md
• MiniGrid neural encoder: docs/experiments/minigrid-neural-unlock.md
• MiniGrid PyTorch DQN lane: docs/experiments/minigrid-torch-lane.md
• Progress: docs/progress/STATUS.md
• Runs: runs/<timestamp>/