{
  "schema": "sys1-research-plan-1",
  "date": "2026-10-25",
  "report": "/blog/embodied-sys1-jev-to-zero-shot",
  "purpose": "Question-conditioned non-language Sys1 between Sys2 and Sys0; zero-shot real grasping is a research target, not an achieved result.",
  "hardware": {
    "robot": "Flexiv; model pending",
    "gripper": "parallel-jaw assumed for pilot; exact device pending",
    "views": "2–3 RealSense RGB views",
    "body_state": true
  },
  "runtime_depth_channel": false,
  "minimal_unit": {
    "outputs": [
      "Choice/Bool including lifecycle and phase",
      "continuous top-down EEF trajectory plus gripper"
    ],
    "data_streams": {
      "T": "Goal-conditioned RGB/body/history to EEF, gripper and essential execution phase/lifecycle labels; common to all action baselines.",
      "B": "Additional paired operation judgments: referent, constraints, evidence, progress and recovery; never supply ground-truth answers as trajectory inputs.",
      "G": "General text/image conditional classification, dynamic candidates and evidence judgments from other scenes."
    },
    "scope": "Single arm, tabletop opaque rigid objects, top-down grasp with yaw, goal revisions and one recovery.",
    "completion": "Frozen zero-shot evaluation on target Flexiv hardware; synthetic or simulation success alone does not complete this unit."
  },
  "model": {
    "status": "proposed",
    "backbone_candidate": "Qwen3.5-9B; exact checkpoint to pin",
    "cost_comparison": "4B",
    "decoder": {
      "layers": 2,
      "dimension": 256,
      "attention_heads": 8
    },
    "context_tokens_initial": [
      4096,
      8192
    ],
    "history_transitions": [
      2,
      4
    ],
    "query_conditioning": "Questions/schema/candidate semantics encoded by backbone; task, EEF and time slot queries decode shared memory.",
    "training_order": [
      "frozen backbone, train decoder/readouts",
      "optional rank-16 LoRA if evidence warrants"
    ],
    "generated_text": false
  },
  "trajectory": {
    "K_initial": 1,
    "H": 8,
    "E": 1,
    "frame": "robot_base",
    "tool": "explicit TCP",
    "position": "absolute xyz metres, normalized using training/predefined domain only",
    "orientation": "Rz(yaw) @ R_down; yaw cos/sin",
    "time_s": [
      0.1,
      0.2,
      0.3,
      0.4,
      0.5,
      0.6,
      0.7,
      0.8
    ],
    "time_origin": "plan_start",
    "gripper": "physical width or explicit open/close event",
    "schema": "eef_topdown_v1; proposed, not the synthetic prototype schema"
  },
  "training_proposal": {
    "T_B_G_sampling": [
      0.6,
      0.25,
      0.15
    ],
    "head_lr": 0.0001,
    "lora_lr": 0.00001,
    "effective_batch": 32,
    "clip": 1,
    "pilot_steps": 50,
    "optional_anchor": "Same-support frozen native option distribution or trained stage-1 reference, never random head. Start lambda_K=0.",
    "compute": "1–2 GPU pilot, then at most 16 A100 80GB for this plan; no training launched."
  },
  "data_proposal": {
    "asset_families_train_dev_test": [
      40,
      10,
      20
    ],
    "episodes_initial": [
      1000,
      3000
    ],
    "decision_points_per_episode": [
      10,
      30
    ],
    "split_unit": "source episode and asset family; hold out camera rigs and semantic combinations separately",
    "teacher_privileged_state": "labels only; never a G0/G1 student input",
    "conditions": {
      "S-sim": "simulation embodied data plus general data",
      "S-mix": "S-sim plus curated public real robot data; no target Flexiv task data"
    },
    "B_questions_per_decision": [
      2,
      4
    ],
    "G_pilot_train_dev_test": [
      2000,
      500,
      500
    ],
    "G_split_unit": "Source image/document, before paraphrase and candidate generation"
  },
  "geometry_conditions": {
    "G0": "RGB/body/history, no explicit camera intrinsics/extrinsics or calibration-derived grids",
    "G1": "G0 plus static camera calibration",
    "G2": "G1 plus RGB-inferred geometry; future extension"
  },
  "simulation_proposal": {
    "groups": [
      "O",
      "N",
      "T",
      "T+G",
      "T+B+G"
    ],
    "first_groups": [
      "O",
      "T",
      "T+G",
      "T+B+G"
    ],
    "rollouts_per_model_approx": 400,
    "training_seeds_main": 3,
    "fairness": "Equal robot episode exposure and control labels; report extra compute and add fixed-token comparison.",
    "metrics": [
      "closed-loop success",
      "paired semantic correctness",
      "constraint satisfaction",
      "recovery",
      "general classification retention",
      "latency",
      "abstention and coverage"
    ]
  },
  "real_proposal": {
    "conditions_per_system": 180,
    "objects": 30,
    "layouts": 3,
    "instruction_conditions": 2,
    "retries_max": 1,
    "lift_height_m": 0.05,
    "hold_s": 2,
    "core_groups": [
      "T+B+G",
      "T",
      "GPT-6 Astra exact snapshot to pin"
    ],
    "pilot_if_limited": 60,
    "claim": "Zero-shot instance/scene transfer, not an unlearned grasp skill.",
    "forbidden_tuning": "No target test task/objects demonstrations, prompt tuning, threshold or checkpoint selection.",
    "allowed_setup": "TCP/body interface and empty workspace tests; static calibration in G1.",
    "matched_conditions": [
      "RGB",
      "body",
      "geometry",
      "history",
      "action interface",
      "Sys0",
      "observation/action budget"
    ],
    "astra_tracks": [
      "Direct versus Agentic separately",
      "capability budget versus wall-clock budget separately"
    ],
    "noninferiority_margin_proposed_pp": 5,
    "sample_size": "Power analysis from pilot required; 180 may be insufficient; clustered paired CI."
  },
  "status": {
    "report": "written",
    "numeric_head": "existing synthetic-only smoke; not a RGB policy",
    "rgb_model": "not_implemented",
    "corpus": "not_built",
    "simulation_experiments": "not_run",
    "real_zero_shot": "not_run",
    "astra_comparison": "not_run"
  }
}