Course navigation
Week 5: VLA ArchitecturesDay 30
π0 / π0.5 / π0.6 / π0.7 deep dive
This is a valid v1.0 placeholder page for the later curriculum arc. Full interactive lab treatment ships after Week 1 dogfooding.
LECTURE & READING
Glossary primer (12 min)
- π0 — Physical Intelligence's first production Modern Robot LearningVision-Language-Action model (VLA)A model that takes images and language as input and outputs robot actions. (Oct 2024). 3.3B params, PaliGemma backbone, flow matching Core ConceptsActionA command the robot sends to its motors, controller, or low-level system. head.
- Flow matching — Generative modeling technique. Train a vector field that transports Data, Distributions & Training IssuesNoiseUnwanted variation or randomness in sensor readings or actuation. to data. Robot LearningInferenceUsing a trained model to make predictions or choose actions.: integrate ODE from Data, Distributions & Training IssuesNoiseUnwanted variation or randomness in sensor readings or actuation.. Cousin of diffusion, often faster.
- π0.5 — Apr 2025. Hierarchical: high-level "Modern Robot LearningSkillA reusable behavior like grasp, push, place, or open drawer." picker + low-level π0 executor.
- π0.6 — Late 2025. DINOv3 vision encoder swap; better Modern Robot LearningGeneralizationThe robot’s ability to work in new situations it has not seen before..
- π0.7 — Apr 16, 2026. JEPA auxiliary loss for Robot LearningRepresentation learningLearning useful internal features automatically from data.. Most generalist π model.
- Open-source weights: π0 base weights released; some fine-tuned variants kept closed.
Real-world analogy
π0 is the "high-end consumer car" of VLAs: doesn't have the raw horsepower of GR00T or the brand of Gemini, but it works reliably across hundreds of tasks. The 0.5/0.6/0.7 versions are model-year refreshes — same chassis, better engine and ECU.
Hour 1 — Reading
- π0 paper, full sections 1–4 (~30 min): https://www.physicalintelligence.company/blog/pi0
- π0.5 blog post (~15 min): https://www.physicalintelligence.company/blog/pi05
- π0.7 blog post (~20 min, Apr 2026): https://www.physicalintelligence.company/blog/pi07
Hour 2 — Read π0's openpi codebase
git clone https://github.com/Physical-Intelligence/openpi
cd openpi
# Read in this order (skim, ~30 min total):
# - src/openpi/policies/pi0/pi0.py (the main architecture)
# - src/openpi/training/data_loader.py (data formatting)
# - src/openpi/training/configs.py (training hyperparameters)LAB
Hour 3 — Lab: run π0 zero-shot on LIBERO + LoRA fine-tune (90 min)
Step 1 — Install openpi (10 min)
cd ~/robo47-il
git clone https://github.com/Physical-Intelligence/openpi
cd openpi
uv pip install -e .Step 2 — Zero-shot inference (20 min)
python -m openpi.scripts.eval_pi0 \
--checkpoint pi0/base \
--task libero_spatial \
--n_episodes 50 \
--output runs/pi0_zsFull source continues in the committed curriculum files. The v1.0 page exposes the day flow and lab surface without inventing content.
Completion controls unlock when this day graduates from placeholder to full lab.