Update 150case_data_enhance_blender_franka_v2/README.md (cup)

150case_data_enhance_blender_franka_v2/README.md

@@ -0,0 +1,55 @@
+# FR3 Duo v2: stack the cups, stack the blocks, put the block in the drawer
+
+Replaces `150case_data_enhance_blender_franka_v1`. LeRobot v2.1, one folder per task, 100 physically admitted episodes each, 30 Hz. Robot: Franka FR3 Duo (two FR3v2 arms on the official Duo mount, a Robotiq 2F-85 on each flange on its 17 mm coupling, from the robot's factory URDF), simulated in MuJoCo 3.3.7 and rendered in a Blender reconstruction of the user's lab site. No DART. Every folder was validated through the official LeRobot v2.1 reader.
+
+| Folder | Task string | Status |
+|---|---|---|
+| `cup/` | `Stack_the_cups_together.` | released |
+| `block/` | `Stack the red cube on top of the green cube.` | in preparation (re-render in the S1 green and red) |
+| `draw/` | `Put_the_block_in_the_drawer.` | in preparation (S1 cabinet appearance; cabinet and block 10 cm farther from the robot) |
+
+The task strings are the S1 datasets' own.
+
+Per frame:
+- `observation.state` (16): measured joint positions (rad) of each arm's 7 joints and its gripper pad opening (m), left arm then right arm;
+- `action` (16): the command executed until the next frame, same layout (absolute joint targets in rad, gripper opening target in m); the last frame repeats the last command;
+- `observation.images.cam_high`: RealSense D455 on the arch behind the robot, pose fitted to the user's camera frame;
+- `observation.images.cam_left_wrist`, `observation.images.cam_right_wrist`: RealSense D405 on the nominal 2F-85 bracket mount (not hardware-calibrated);
+- all three views are 4:3 (224x168) black-padded to 224x224, H.264.
+
+## Scene (all tasks)
+
+- **Desk.** The user's desk (1.2 x 0.6 m under a dark cloth), moved 15 cm toward the robot from its measured place: its near edge is 0.522 m in front of the robot's mounting point, and the objects sit 0.24-0.26 m behind that edge.
+- **Objects.** Copied from the S1 task scenes (their MuJoCo bodies verbatim), placed on the FR3 desk.
+- **Start pose.** The robot's start pose as read from the real robot (FCI), not perturbed.
+- **Rendering.** Blender 4.5.14 LTS Cycles, 12 samples with denoising. Each episode gets its own seeded S1-style lighting and exposure randomisation.
+
+## Admission (every released episode)
+
+- the frozen RSSR task contract and the hard constraints, including joint speed <= 2 rad/s and tracking error <= 0.3 rad;
+- a bit-exact independent replay;
+- one grasp per object, no state-action conflict, no idle segment;
+- >= 6 cm from the camera arch in every frame, measured exactly (dense samples on every robot collision surface against the arch tubes);
+- stacking tasks: >= 10 cm between the two arms in every frame, measured the same way;
+- within the FR3 v2 position-based velocity limits at 1x, and >= 0.05 rad inside the first-generation FR3 joint ranges.
+
+## Generation
+
+- **Sim2.** A rule controller (no LLM call) per task.
+- **Refine.** The same placement is run again with refine rules, most aggressive first; the first setting that passes full admission is kept:
+  - R1 no detour: the upper hand carries its object straight to a point beside the stack;
+  - R2 overlap: both arms move at once wherever their joint paths are clear together;
+  - R3 speed: line and joint speeds x1.6 (joints <= 0.72 rad/s, <= 0.5 rad/s while holding an object), 0.4 s gripper actions;
+  - R4 short exit: after the last release the hand only rises.
+- **Selection.** The first 100 refined episodes in placement order.
+
+## cup/: stack the cups
+
+- **Objects.** Two identical tapered paper cups, 8 cm across the top and 9 cm high, 32 wall facets (S1's measured cup).
+- **Placement augmentation (as S1).** Each cup in a 15 cm disk around its nominal position with yaw in U(-15, 15) deg; the stack region in a 5 cm disk.
+- **Counts.** Sim2: 209 of 220 placements admitted (median 29.6 s for the released ones). Refine: 206 of 209. Released: the first 100 in placement order, 85 with all rules at x1.6, 8 at x1.3, 7 with the upper hand holding its cup at the lift pose.
+- **Episodes.** 47,957 frames; 12.9 / 15.3 / 22.5 s (min / median / max). Closest approach to the arch 6.52 cm; closest approach between the arms 10.53 cm.
+
+## Notes
+
+- `meta/source_cases.jsonl` lists, for each episode: its placement, run, record and episode hashes, refine setting, render receipt and video hashes. Its `blender` field is empty (a packaging key error); every episode was rendered with Blender 4.5.14 LTS.