rooty2020/r2s-straw-v2n-rollouts

r2s straw v2n: pi0.5 student rollouts (sim2sim, LeRobot v2.1, no images)

Closed-loop rollouts of the pi0.5 student of each round, in the sim worlds and in the hidden truth world.

kind	what	world
collect	round r's start: the previous student (policy_round r-1) on the 60 training scenes. Its failures drive the diagnosis	world_{r-1}
eval_new_world	the round-r student on dev24 + D96 in the round's sim world	world_r
eval_truth	the round-r student on the same scenes in the hidden truth world (what counts as real success)	truth
eval_old_world	the round-r student in the previous world (dev24 only)	world_{r-1}
eval_prev_new	the previous student in the new world	world_r

Same scene + round across kinds = the same initial scene (dev24: sets S1/S2, D96: set D).

1572 episodes, 1028652 frames.

round	kind	world	policy_round	episodes	success
0	eval_new_world	world_0	0	120	6
0	eval_truth	truth	0	120	1
1	collect	world_0	0	60	2
1	eval_new_world	world_1	1	120	8
1	eval_old_world	world_0	1	24	3
1	eval_prev_new	world_1	0	120	3
1	eval_truth	truth	1	120	5
2	collect	world_1	1	60	3
2	eval_new_world	world_2	2	120	11
2	eval_old_world	world_1	2	24	4
2	eval_prev_new	world_2	1	120	9
2	eval_truth	truth	2	120	8
3	collect	world_2	2	60	1
3	eval_new_world	world_3	3	120	8
3	eval_old_world	world_2	3	24	2
3	eval_prev_new	world_3	2	120	5
3	eval_truth	truth	3	120	9

Setup

• Task pnp_straw: a Franka FR3 picks up a straw and inserts it into the hole of a cup's lid. Prompt: "pick up the straw and insert it into the hole of the cup's lid".
• Run straw_s2s_v2n_s0 of the r2s evolve loop: a sim world, a teacher and a pi0.5 student (LoRA) evolve round by round. Each round repairs the sim's physics, generates new teacher demos, fine-tunes the student (1500 steps, batch 16, from the previous round's checkpoint) and evaluates it.
• sim2sim: a hidden truth world (another MuJoCo physics setting) plays the role of reality. The loop never reads its parameters. It only sees replays and evaluations run in it.
• Simulator: MuJoCo, control period 0.11 s. No images are included (state, action and object state only).

Worlds (meta/r2s_worlds.json)

world_r is the sim world after round r's sim repair. The round-r student was trained and evaluated in it. world_0 is the base world. The loop evolves only these four parameters; all other physics equals the base world. The table shows full values (bold = differs from the base world). meta/r2s_worlds.json stores only the changed values (physics_delta).

world	objects.cup.mass	robot.grip_force	robot.pad_rolling	robot.pad_torsion
truth	0.025	95	0.0001	0.002
world_0	0.025	130	0.0001	0.005
world_1	0.065	76	0.0001	0.0023
world_2	0.065	100	0.0003	0.002
world_3	0.055	95.5	0.0001296	0.002

Per-frame features

key	shape	meaning
observation.state	8	measured joint positions q1..q7 (rad) and gripper opening
action	8	joint position goals (q1..q7) and the gripper command sent to the controller during this frame
observation.environment_state	14	straw pose then cup pose: xyz (m) + quaternion (w, x, y, z), world frame
observation.sim_time	1	simulator time (s); frames are 0.11 s apart

timestamp = frame_index / 9 because LeRobot stores an integer fps. The true period is 0.11 s (use observation.sim_time).

Per-episode metadata (meta/r2s_episodes.jsonl, one line per episode_index)

• round, kind, set (dev24:S1, dev24:S2, D96:D, train), policy_round: which round's student
• world, physics_delta: the world the rollout was simulated in
• success, failure_key (not_lifted, not_vertical, not_inserted, ...), reason, stages
• scene, variant, seed, goal_xy, episode_s, n_frames, source
• The last simulator tick (the state after the final action) is dropped so that every frame has an action.