zrgong/rmbench-plus-plus-press-button-noop

RMBench++ Press Button No-op v2.0

This dataset contains 100 RT-rendered scripted-expert demonstrations for the RMBench++ press_button_noop v2 training split. Only the no-op dataset was regenerated for v2; the published vanish and update training datasets are unchanged.

Each number card gives the required registered-press total for the counting button below it. The left arm operates both counting buttons. For every episode, the frozen seed deterministically samples K distinct logical press slots, with K in {1, 2, 3}. The first physical attempt at each sampled slot is not registered, so the corresponding acknowledgement lamp stays dark and the policy must retry. The right arm confirms once after both registered totals are correct. The first registered confirm edge is irreversible.

Canonical task prompt

Each number card shows how many successful presses its counting button requires. Use the left arm for both counting buttons. Watch the number cards and indicator lights throughout the task, and count only presses acknowledged by the corresponding light. When both successful-press totals match the current targets, confirm once with the right arm.

The prompt is identical for training and evaluation and does not disclose the event schedule. Training uses seeds 1 through 100; held-out evaluation uses seeds 100001 through 100100. Task parameters are deterministic for a frozen seed; there are no online no-op coin flips.

No-op and lamp contract

The no-op schedule is sampled without replacement from feasible logical slots (button, required_press_index). All scheduled no-ops must occur before confirm can succeed. Once a press is registered, its lamp intensity is computed only from that button's current joint position:

intensity = clip(max(0, -qpos) / 0.005, 0, 1)

A no-op keeps the acknowledgement gate closed, so its lamp remains dark at the same physical depression. A released button is dark. Robot pose, arm identity, action values, trajectory, and gripper state are not inputs to the lamp.

Dataset contents

• 100 episodes using the frozen training seeds 1 through 100.
• 62,387 frames at 30 FPS.
• One Parquet file per episode and one H.264 MP4 per episode per camera.
• Native RMBench 14D observation.state and 14D action: six left-arm joints, left gripper, six right-arm joints, and right gripper.
• Policy views: head_camera, left_camera, and right_camera, each RGB 320 x 240. front_camera is included for QA and visualization and is not a policy input under the v2 evaluation protocol.
• Per-frame episode seed, canonical-prompt provenance, and JSON-encoded rmbench_plus audit metadata. Seed and audit fields are labels and must not be exposed to a policy.

The embedded meta/rmbench_plus_manifest.json records protocol version rmbench-plus-v2 and the disjoint train/eval seed ranges. This repository contains training demonstrations only, not held-out evaluation rollouts.

Frozen no-op distribution

split	K=1	K=2	K=3	left only	middle only	both sides
train	27	35	38	25	30	45
eval	33	32	35	28	32	40

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="zrgong/rmbench-plus-plus-press-button-noop",
    video_backend="pyav",
)

Validation and provenance

The uploaded tree passed the RMBench++ LeRobot v3 validator with:

• codebase_version == "v3.0"
• 100 episode metadata rows
• 100 data Parquet files
• 400 video files across four camera keys
• 14D state and action reads through LeRobotDataset
• exact frozen seeds 1 through 100
• all scheduled no-op events observed and all 100 scripted-expert confirmations successful, with no protocol-failure events

The protocol, simulator, converter, validator, dashboard, and manifest are pinned at RMBench++ code commit ca2d2c56b0476c46c3c3807ab5057eb28492ddb6.

Tags v1.0.0 and v1.1.0 preserve the previous exactly-one-no-op releases. Tag v2.0.0 identifies this multi-no-op regeneration.

These demonstrations are task implementation and scripted-expert data evidence. They are not closed-loop model-evaluation results. RMBench++ v2 is a pilot protocol, not a general-memory leaderboard.