"""Your personal reading list, linked to landscape branches / subcategories.

Each paper attaches to ONE node:
  - a subcategory  (a method node id, e.g. "diffusion-policy", "clip", "moe-vlm")
  - or a branch    (a family key,      e.g. "BC", "Contrastive", "VLA")

It then shows in that node's detail panel under "📌 My reading", and a gold
badge with the count appears on the map (on the node, and summed on its family
hub). Works for BOTH the robot and the VLM landscapes (matched by id).

Send me your papers and I'll classify each into the right node and fill this in.

Format:
    dict(node="<id or family key>", title="...", url="...", note="..."),
url and note are optional.

Node ids you can use:
  ROBOT methods: flow-matching-policy, diffusion-policy, tokenized-bc, energy-based-bc,
    value-based-rl, policy-gradient-rl, off-policy-ac, offline-rl, maxent-irl, gail,
    forward-dynamics-mpc, latent-imagination, generative-video-wm, action-conditioned-wm,
    world-action-model, occupancy-latent-wm, decision-transformer, trajectory-diffusion,
    goal-conditioned, hrl, meta-learning, llm-planner, vlm-affordance, pid-control, lqr,
    classical-mpc, motion-planning, vla-foundation, vla-rl, domain-randomization,
    sim2real-adapt, visual-pretrain, latent-action
  ROBOT branches: BC, Reinforcement, "Offline RL", "Inverse RL", "Model-Based", Sequence,
    "Goal-Cond.", Hierarchical, "Meta-Learning", "LLM-Orchestration", Classical, VLA,
    Sim2Real, Representation
  VLM methods: clip, siglip, coca, blip, git, flamingo, blip2, llava, fuyu, chameleon,
    frontier, glip, kosmos2, llavanext, qwenvl, videollava, donut, vit, ssl, instructtune,
    mmrlhf, unified, flava, beit3, small-vlm, moe-vlm, agentic-vlm, mm-rag
  VLM branches: Contrastive, Generative, Bridge, Native, Grounding, AnyRes, Video,
    Document, Backbone, Posttrain, Unified, Masked, Efficient, MoE, Agentic
"""

MY_PAPERS = [
    dict(node="Video", title="Cambrian-P: Pose-Grounded Video Understanding (2026)",
         url="https://arxiv.org/abs/2605.22819",
         note="video MLLM + per-frame camera-pose tokens → better spatial reasoning"),
    dict(node="vla-foundation", title="Robotic Manipulation is Vision-to-Geometry Mapping — VGA (2026)",
         url="https://arxiv.org/abs/2604.12908",
         note="conditions action on pretrained native 3D geometry, not a VLM/video backbone"),
    # ---- send more papers and I'll classify each into a node ----
]