"""Generate a separate 'Evolution Timeline' page for robot learning.
A time-ordered lineage tree: a central time spine (the linear trunk), with each
method/milestone placed at its year and arrows showing what it was developed
FROM (influence / "developed after"). Hovering a node traces its full lineage
(ancestors above, descendants below). Self-contained HTML; same dark theme as
the landscape.
Run: .venv_robot_paradigms/bin/python gen_evolution.py -> robot_evolution.html
"""
import json
import html as html_lib
import robot_paradigms_app as app
import gen_landscape
FAMILY_LABEL = {
"Classical": "Classical Control", "BC": "Behavioral Cloning",
"Reinforcement": "Reinforcement Learning", "Offline RL": "Offline RL",
"Inverse RL": "Inverse RL", "Model-Based": "World Models",
"Sequence": "Sequence Models", "Goal-Cond.": "Goal-Conditioned",
"Hierarchical": "Hierarchical", "Meta-Learning": "Meta-Learning",
"LLM-Orchestration": "LLM / VLM",
}
FAMILY_COLOR = {k: app.FAMILY[k][0] for k in app.FAMILY}
FAMILY_COLOR["Classical"] = "#64748b"
# (id, name, year, family, [parent ids], note) — parents = "developed from / after"
TIMELINE = [
("pid", "PID control", 1922, "Classical", [], "Proportional-integral-derivative feedback — the oldest controller."),
("dp", "Dynamic Programming",1957, "Classical", [], "Bellman's value recursion — the root of value-based RL."),
("lqr", "LQR (Kalman)", 1960, "Classical", ["dp"], "Optimal linear feedback; a classical-control precursor (RL's roots are dynamic programming)."),
("astar", "A* search", 1968, "Classical", [], "Optimal graph search — classic motion planning."),
("mpc", "MPC", 1978, "Classical", ["lqr"], "Receding-horizon optimization with a known model."),
("qlearning", "Q-Learning", 1989, "Reinforcement", ["dp"], "Learn action-values from rewards; act greedily."),
("alvinn", "ALVINN", 1989, "BC", [], "First neural behavioral-cloning driver."),
("reinforce", "REINFORCE", 1992, "Reinforcement", [], "The original policy-gradient estimator."),
("prm", "PRM / RRT", 1996, "Classical", ["astar"], "Sampling-based motion planning in continuous spaces."),
("options", "Options (HRL)", 1999, "Hierarchical", ["qlearning"], "Temporal abstraction: sub-policies over time."),
("irl", "Inverse RL", 2000, "Inverse RL", ["qlearning"], "Recover the reward that explains expert behavior."),
("maxentirl", "MaxEnt IRL", 2008, "Inverse RL", ["irl"], "Probabilistic IRL via maximum entropy."),
("dqn", "DQN", 2013, "Reinforcement", ["qlearning"], "Deep Q-Networks crack Atari from pixels."),
("trpo", "TRPO", 2015, "Reinforcement", ["reinforce"], "Trust-region policy optimization."),
("ddpg", "DDPG", 2015, "Reinforcement", ["dqn", "reinforce"], "Off-policy actor-critic for continuous control."),
("a3c", "A3C", 2016, "Reinforcement", ["dqn", "reinforce"], "Asynchronous advantage actor-critic."),
("gail", "GAIL", 2016, "Inverse RL", ["maxentirl", "trpo"], "Adversarial imitation — match the expert via a discriminator."),
("rl2", "RL²", 2016, "Meta-Learning", ["a3c"], "An RNN that learns a learning algorithm."),
("ppo", "PPO", 2017, "Reinforcement", ["trpo"], "Clipped policy gradient — the RL workhorse."),
("maml", "MAML", 2017, "Meta-Learning", ["reinforce"], "Learn an init that adapts in a few gradient steps."),
("her", "HER", 2017, "Goal-Cond.", ["ddpg"], "Relabel failures as successes for goal-reaching."),
("optioncritic", "Option-Critic", 2017, "Hierarchical", ["options", "a3c"], "Learn options end-to-end."),
("worldmodels", "World Models", 2018, "Model-Based", ["dqn"], "Train a policy inside a learned generative world."),
("sac", "SAC", 2018, "Reinforcement", ["ddpg"], "Maximum-entropy off-policy actor-critic."),
("td3", "TD3", 2018, "Reinforcement", ["ddpg"], "Twin-critic fix for overestimation."),
("dreamer", "Dreamer", 2019, "Model-Based", ["worldmodels"], "Learn behaviors by latent imagination."),
("pearl", "PEARL", 2019, "Meta-Learning", ["maml", "sac"], "Off-policy meta-RL with latent task variables."),
("bcq", "BCQ", 2019, "Offline RL", ["ddpg"], "Batch-constrained offline RL."),
("cql", "CQL", 2020, "Offline RL", ["bcq", "sac"], "Conservative Q-learning for offline data."),
("muzero", "MuZero", 2020, "Model-Based", ["dreamer", "dqn"], "Plan with a learned model, no rules given."),
("gpt3", "GPT-3 (LLMs)", 2020, "LLM-Orchestration", [], "Large language models — an external force on robotics."),
("dt", "Decision Transformer",2021, "Sequence", ["cql", "gpt3"], "Cast RL as return-conditioned sequence modeling."),
("implicitbc", "Implicit BC", 2021, "BC", ["alvinn"], "Energy-based behavioral cloning."),
("diffuser", "Diffuser", 2022, "Sequence", ["dt"], "Plan by denoising whole trajectories."),
("rt1", "RT-1", 2022, "BC", ["alvinn", "gpt3"], "Tokenized robot transformer at scale."),
("gato", "Gato", 2022, "BC", ["dt"], "One tokenized transformer for many tasks."),
("saycan", "SayCan", 2022, "LLM-Orchestration", ["gpt3"], "LLM proposes, affordances ground."),
("cap", "Code as Policies", 2022, "LLM-Orchestration", ["gpt3"], "LLM writes robot control code."),
("tdmpc", "TD-MPC", 2022, "Model-Based", ["dreamer", "muzero"], "Latent planning + a learned value."),
("diffusionpolicy","Diffusion Policy", 2023, "BC", ["implicitbc", "diffuser"], "Generate actions by denoising — multi-modal BC."),
("aloha", "ACT / ALOHA", 2023, "BC", ["alvinn"], "Action chunking for fine bimanual teleop."),
("rt2", "RT-2 (VLA)", 2023, "BC", ["rt1", "gpt3"], "Vision-language-action model from a VLM."),
("flow", "Flow Matching", 2023, "BC", ["diffusionpolicy"], "Straight-line generative action head."),
("dreamerv3", "DreamerV3", 2023, "Model-Based", ["dreamer"], "One world-model recipe across many domains."),
("voxposer", "VoxPoser", 2023, "LLM-Orchestration", ["saycan", "cap"], "VLM-composed 3D value maps for manipulation."),
("unipi", "UniPi", 2023, "Model-Based", ["worldmodels"], "Generate video plans, then act."),
("openvla", "OpenVLA", 2024, "BC", ["rt2"], "Open vision-language-action model."),
("pi0", "π₀ (flow VLA)", 2024, "BC", ["flow", "openvla"], "Flow-matching action head on a VLM trunk."),
("rekep", "ReKep", 2024, "LLM-Orchestration", ["voxposer"], "Relational keypoint constraints from a VLM."),
("groot", "GR00T N1", 2024, "Model-Based", ["rt2", "diffusionpolicy"], "Humanoid foundation policy."),
("vjepa2", "V-JEPA 2", 2025, "Model-Based", ["dreamerv3"], "Action-conditioned latent world model + planning."),
("wam", "World-Action Models",2025, "Model-Based", ["unipi", "pi0"], "Imagine the future video AND the actions."),
("pi05", "π₀.₅ / π₀.₆", 2025, "BC", ["pi0"], "Newer flow VLAs with broader generalization."),
]
# Timeline node -> reuse a verified landscape "Learn" link (same concept)
REUSE = {
"pid": "pid-control", "lqr": "lqr", "mpc": "classical-mpc",
"astar": "motion-planning", "prm": "motion-planning",
"qlearning": "value-based-rl", "dqn": "value-based-rl",
"trpo": "policy-gradient-rl", "ppo": "policy-gradient-rl", "sac": "off-policy-ac",
"options": "hrl", "optioncritic": "hrl", "irl": "maxent-irl", "maxentirl": "maxent-irl",
"gail": "gail", "rl2": "meta-learning", "maml": "meta-learning", "pearl": "meta-learning",
"her": "goal-conditioned", "worldmodels": "generative-video-wm",
"dreamer": "latent-imagination", "dreamerv3": "latent-imagination",
"bcq": "offline-rl", "cql": "offline-rl", "dt": "decision-transformer",
"diffuser": "trajectory-diffusion", "implicitbc": "energy-based-bc",
"diffusionpolicy": "diffusion-policy", "flow": "flow-matching-policy",
"cap": "llm-planner", "voxposer": "vlm-affordance", "openvla": "tokenized-bc",
"pi0": "flow-matching-policy", "vjepa2": "action-conditioned-wm", "wam": "world-action-model",
}
# Timeline-specific explainers (verified reachable June 2026)
EXTRA = {
"dp": ("Dynamic Programming & Value Iteration — Brunton (video)", "https://www.youtube.com/watch?v=sJIFUTITfBc"),
"alvinn": ("Behavioral Cloning for Self-Driving — Udacity (video)", "https://www.youtube.com/watch?v=xxuEDx_zlsU"),
"reinforce": ("Policy Gradients & PPO — Arxiv Insights (video)", "https://www.youtube.com/watch?v=5P7I-xPq8u8"),
"a3c": ("Actor-Critics & A3C — AI Summer (illustrated)", "https://theaisummer.com/Actor_critics/"),
"ddpg": ("DDPG Explained (video)", "https://www.youtube.com/watch?v=ClKZaaIJr6Y"),
"td3": ("TD3 Explained (video)", "https://www.youtube.com/watch?v=qXyLPHe2qv8"),
"muzero": ("MuZero — DeepMind", "https://deepmind.google/blog/muzero-mastering-go-chess-shogi-and-atari-without-rules/"),
"tdmpc": ("TD-MPC2 — project page", "https://www.tdmpc2.com/"),
"gpt3": ("How GPT-3 Works — Jay Alammar", "https://jalammar.github.io/how-gpt3-works-visualizations-animations/"),
"gato": ("Gato: A Generalist Agent — DeepMind", "https://deepmind.google/discover/blog/a-generalist-agent/"),
"unipi": ("UniPi — Google Research", "https://research.google/blog/unipi-learning-universal-policies-via-text-guided-video-generation/"),
"aloha": ("ALOHA / ACT — project page", "https://tonyzhaozh.github.io/aloha/"),
"rt1": ("RT-1 — project page", "https://robotics-transformer1.github.io/"),
"rt2": ("RT-2 (VLA) — project page", "https://robotics-transformer2.github.io/"),
"saycan": ("SayCan — project page", "https://say-can.github.io/"),
"rekep": ("ReKep — project page", "https://rekep-robot.github.io/"),
"groot": ("Isaac GR00T N1 — NVIDIA", "https://developer.nvidia.com/isaac/gr00t"),
"pi05": ("π₀.₅ — Physical Intelligence", "https://www.physicalintelligence.company/blog/pi05"),
}
def learn_for(node_id):
if node_id in EXTRA:
return {"title": EXTRA[node_id][0], "url": EXTRA[node_id][1]}
if node_id in REUSE and REUSE[node_id] in gen_landscape.LEARN:
t, u = gen_landscape.LEARN[REUSE[node_id]]
return {"title": t, "url": u}
return None
def build_data():
nodes = [dict(id=i, name=n, year=y, fam=f, parents=ps, note=note, learn=learn_for(i))
for (i, n, y, f, ps, note) in TIMELINE]
fams = [dict(key=k, label=FAMILY_LABEL[k], color=FAMILY_COLOR[k]) for k in FAMILY_LABEL]
return dict(nodes=nodes, families=fams)
TEMPLATE = r"""
Robot Learning — Evolution Timeline
⏳ Robot Learning — Evolution Timeline
— arrows point from an idea to what was built on it. Hover to trace lineage · click ↗ to read its best explainer.
"""
def main():
data = build_data()
html = TEMPLATE.replace("__DATA_JSON__", json.dumps(data))
with open("robot_evolution.html", "w", encoding="utf-8") as fh:
fh.write(html)
print(f"wrote robot_evolution.html ({len(html):,} chars) — {len(data['nodes'])} milestones")
if __name__ == "__main__":
main()