EmbodiedSWE/pc_gpu_ram_franka_osc.raw_15hz_20260907

pcgpuram — 200 demonstrations (assembly.pcgpuram.franka.osc)

Generated by the CoSiGen datagen ladder (`geno50v3`): {'scene': 5, 'strategy': 20, 'phase': 50} -> physics set of 200 episodes, every one replay-verified (200/200): the recorded actions, fed back open-loop into the rebuilt world under the recorded control law, reproduce the task's success. Seed solution: `CoSiGenSolutions/assembly/pcgpuram/franka/osc@8d74c75`.

• —scenes/cells: 5 scene variants, 15 cells (scene/strategy[/phase]); sources under scenes/
• —sim dt 0.004166666666666667 s, decimation 16; controller FrankaRobot with leaves [{'class': 'OperationalSpaceController', 'controlperiod': 16}, {'class': 'JointController', 'controlperiod': 16}]
• —episode length: 570 / 2400 / 2850 control steps (min / median / max)
• —renders: 600/600 items = 3 looks x 200 episodes; cameras ['front', 'high', 'wrist']
• —size: 19.4 GB

Layout

physicsset.json the 200 delivered episodes (paths below); scene/strategy/phaseset.json = the nested rungs episodes.jsonl one row per episode (path, cell, steps, seed, replayverified, physicalparams, videos, ...) rendermanifest.json per (episode, look) render record data/<batch>/epNNNN/ meta.json cell, seed, steps, successstep, controller (the law the episode ran under), controllerchanges, noise, replay verdict traj.npz per-step arrays, one row per control step, state recorded BEFORE the step: ['action', 'robot/controller/0/prevaction', 'robot/jointefforttarget', 'robot/jointpos', 'robot/jointpostarget', 'robot/jointvel', 'robot/root', 'scene/card', 'scene/case', 'scene/graspheld', 'scene/grasprelp', 'scene/grasprelq', 'scene/ram0', 'scene/ram1'] action (8-dim) is the commanded action at that step imgs/<camera>.mp4, <camera>draw1.mp4, <camera>draw2.mp4 the episode rendered under each look imgs/render<camera>[drawN].json render parameters of that video

Reading an episode

import json, numpy as np meta = json.load(open("data/<batch>/ep0000/meta.json")) tr = dict(np.load("data/<batch>/ep0000/traj.npz")) tr["action"].shape, tr["robot/jointpos"].shape # (T, 8), (T, njoints)