"""Generate an experimental animated story for the robot-learning landscape.
This is intentionally separate from the deployed Gradio wrapper. It is a
research/prototype page that turns the existing paper atlas into a smoother
"how the field got here" narrative.
Run from this directory:
.venv_robot_paradigms/bin/python gen_unified_story.py
Out:
robot_unified_story.html
"""
from __future__ import annotations
import json
from pathlib import Path
import robot_paradigms_app as app
ERAS = [
{
"id": "control",
"years": "1922-1988",
"title": "Control bedrock",
"claim": "Robots start as engineered feedback systems: stabilize, track, avoid collisions, optimize a short horizon.",
"nodes": ["PID", "LQR", "A*", "MPC", "Trajectory optimization"],
"takeaway": "The runtime object is a controller or planner, not a learned policy.",
},
{
"id": "rl",
"years": "1989-2017",
"title": "Learning objectives split",
"claim": "Imitation, RL, inverse RL, goal-conditioned RL, hierarchy, and meta-learning become separate ways to train behavior.",
"nodes": ["BC", "Q-learning", "Policy gradient", "Actor-critic", "IRL/GAIL", "HER/HRL/MAML"],
"takeaway": "The clean distinction is objective, not architecture.",
},
{
"id": "deep",
"years": "2018-2021",
"title": "Offline and sequence views",
"claim": "Replay buffers and logged trajectories become datasets; transformers recast control as conditional sequence generation.",
"nodes": ["SAC/TD3", "CQL/IQL/AWAC", "Decision Transformer", "Trajectory Transformer", "Gato"],
"takeaway": "BC, offline RL, and sequence models start to overlap.",
},
{
"id": "gen",
"years": "2022-2024",
"title": "Generative action heads",
"claim": "Action chunking, diffusion, flow matching, and tokenized actions replace single-mode MSE regression.",
"nodes": ["ACT/ALOHA", "Diffusion Policy", "Flow Matching", "Action tokens", "Octo/OpenVLA"],
"takeaway": "Modern VLA policies mostly differ by action head and data, not by a new objective.",
},
{
"id": "vlm",
"years": "2023-2026",
"title": "VLM/VLA foundation policies",
"claim": "Vision-language pretraining injects semantics; robot data teaches control. Cross-embodiment datasets make transfer plausible.",
"nodes": ["RT-2/RT-X", "Open X-Embodiment", "OpenVLA", "pi0/OpenPI", "Gemini Robotics", "GR00T/Helix/SmolVLA"],
"takeaway": "VLA is a trunk + action interface + data recipe.",
},
{
"id": "wm",
"years": "2018-2026",
"title": "World models return",
"claim": "Learned simulators support planning, imagination, evaluation, data generation, and video/action proposal.",
"nodes": ["World Models", "Dreamer/TD-MPC", "UniPi/Genie", "V-JEPA 2", "AC-WM", "WAM"],
"takeaway": "Separate action-conditioned simulators from world-action proposal models.",
},
{
"id": "unified",
"years": "2025-2026",
"title": "Unified robot model frontier",
"claim": "The frontier is no longer one algorithm. It is a stack: VLA policy, world model, planner, evaluator, safety controller, and data engine.",
"nodes": ["Policy", "Planner", "World model", "Verifier", "Data engine", "Safety layer"],
"takeaway": "The landscape should explain roles and interfaces before listing names.",
},
]
LANES = [
{"id": "control", "label": "Control substrate", "color": "#64748b"},
{"id": "objective", "label": "Learning objective", "color": "#f97316"},
{"id": "action", "label": "Action representation", "color": "#38bdf8"},
{"id": "world", "label": "World model", "color": "#a78bfa"},
{"id": "unified", "label": "Unified stack", "color": "#34d399"},
]
FLOW = [
["control", "control", "PID / LQR / MPC"],
["control", "objective", "known cost -> learned reward/policy"],
["objective", "action", "BC/RL losses need multimodal heads"],
["action", "unified", "VLA policies close the loop"],
["objective", "world", "model-based RL and imagination"],
["world", "unified", "simulator, evaluator, data engine"],
["control", "unified", "MPC/safety still wraps deployment"],
]
UPDATES = [
{
"title": "VLA surveys now treat action representation as the central split",
"body": "Tokenized actions, diffusion, and flow matching are better shown as sibling action heads under BC/foundation-policy training.",
},
{
"title": "World-model robotics needs two branches",
"body": "Action-conditioned world models take actions as inputs and support counterfactual planning; world-action models output a successful future plus actions.",
},
{
"title": "Unified does not mean one box",
"body": "Current systems are hybrid stacks: VLM/VLA semantic trunk, continuous control head, world model, classical planner/MPC, evaluator, and safety guard.",
},
{
"title": "Data infrastructure is becoming a first-class axis",
"body": "Open X-Embodiment, DROID/Bridge/LeRobot-style datasets, synthetic engines, human video, and auto-generated labels explain many model differences.",
},
]
SOURCE_LINKS = [
("RT-2", "https://arxiv.org/abs/2307.15818"),
("Open X-Embodiment / RT-X", "https://arxiv.org/abs/2310.08864"),
("OpenVLA", "https://arxiv.org/abs/2406.09246"),
("pi0", "https://arxiv.org/abs/2410.24164"),
("GR00T N1", "https://arxiv.org/abs/2503.14734"),
("SmolVLA", "https://arxiv.org/abs/2506.01844"),
("VLA systematic review", "https://arxiv.org/abs/2507.10672"),
("VLA real-world review", "https://arxiv.org/abs/2510.07077"),
("World-model robotics survey", "https://arxiv.org/abs/2605.00080"),
("VLA data engines survey", "https://arxiv.org/abs/2604.23001"),
("Foundation models in robotics review", "https://arxiv.org/abs/2604.15395"),
]
def paper_rows():
rows = []
for idx, (title, authors, year, tag) in enumerate(app.PAPERS, start=1):
rows.append({"i": idx, "title": title, "authors": authors, "year": year, "tag": tag})
return rows
def build_data():
rows = paper_rows()
tags = sorted({r["tag"] for r in rows})
return {
"eras": ERAS,
"lanes": LANES,
"flow": FLOW,
"updates": UPDATES,
"papers": rows,
"tags": tags,
"sources": [{"title": t, "url": u} for t, u in SOURCE_LINKS],
"counts": {"papers": len(rows), "tags": len(tags)},
}
TEMPLATE = r"""
Robot Learning Landscape - Animated Unified Story
Robot Learning Landscape - Animated Unified Story
Landscape Merge Recommendations
"""
def render() -> str:
return TEMPLATE.replace("__DATA_JSON__", json.dumps(build_data(), ensure_ascii=False))
def main() -> None:
Path("robot_unified_story.html").write_text(render(), encoding="utf-8")
print("Wrote robot_unified_story.html")
if __name__ == "__main__":
main()