zhb10086/umi-toybench-dual

UMI ToyBench Dual

中文说明

umi-toybench-dual 是一个基于 PyBullet 的合成长时序操作基准，用于测试双夹爪、 语言条件策略的组合泛化能力。场景中包含 9 个颜色和尺寸不同的方块以及 3 个目标区域； 左右夹爪分别用一对无碰撞小球显示，因此从全局、左腕和右腕三路相机中都能直接观察 夹爪位置与开合状态。

本数据集适合用于：

• 验证机器人学习数据管线是否正确连通；
• 训练行为克隆、ACT、视觉—语言—动作等策略；
• 在 ID、组合、语言改写、语义组合和视觉变化下进行受控泛化评测；
• 分析长时序任务分解、动作分块和语义瓶颈。

与本数据集配套的两个 Hugging Face 数据包为：

数据集	作用
umi-toybench-dual	完整低维 benchmark，以及训练/验证集 RGB 数据
umi-toybench-dual-1000	1,000 条 episode 实验子集、ACT 训练包和带时间范围的 latent-skill 标注

本仓库不镜像 LIBERO，不包含 checkpoint、优化器状态、运行日志、失败实验或评测输出。

数据规模

数据包	Episode 数	动作或帧数	用途
raw/	12,900	3,095,162 个动作	低维轨迹、任务/场景定义、语义片段及 horizon-16 样本
rgb_train_validation/	6,900	1,417,514 个观测帧	训练/验证 RGB、signals 和 episode metadata

低维数据包含 64,100 个已验证语义片段，以及分布在 202 个 shard 中的 798,067 个 stride-4、horizon-16 样本。RGB 数据包含 34,500 个 episode 文件，总大小约 6.57 GB。

数据划分与受控分布变化

每个任务定义包含 50 个物理场景副本。训练集包含 132 个任务定义，验证集包含 6 个， 每类测试集包含 24 个。

Split	任务定义数	Episode 数	变化因素
train	132	6,600	36 个原子任务、84 个已见有序组合、12 个多解任务
validation	6	300	已见任务程序，使用新的 ID 场景
test_id	24	1,200	已见任务程序，使用新的 ID 场景
test_composition	24	1,200	未见的有序任务组合，但各原子任务均在训练中出现
test_paraphrase	24	1,200	已见任务程序，使用训练中未出现的语言改写
test_semantic	24	1,200	由已见原子任务构成的未见三段式语义宏
test_visual	24	1,200	已见任务程序，使用未见渲染域

在适用的测试族之间，数据保留了匹配的结构化场景组，因此可以在场景结构不变时， 单独比较组合顺序、语言、语义抽象或渲染变化带来的影响。

只有训练集和验证集提供预渲染 RGB。测试集仅发布低维轨迹；正式闭环评测应从冻结的 场景与任务记录重建环境并在线渲染，避免训练或模型选择过程意外读取测试图像。

目录结构

.
├── COMPLETE.json
├── metadata.json
├── RELEASE.json
├── raw/
│   ├── COMPLETE.json
│   ├── metadata.json
│   ├── protocol.json
│   ├── run_config.json
│   ├── task_catalog.json
│   ├── audit.json
│   └── shards/shard_*/
│       ├── episodes.jsonl
│       ├── segment_labels.jsonl
│       ├── samples.parquet
│       ├── trajectories/<split>/*.npz
│       ├── manifest.json
│       └── audit.json
└── rgb_train_validation/
    ├── episodes.jsonl
    ├── episodes/<prefix>/<episode_key>/
    │   ├── global.mp4
    │   ├── left.mp4
    │   ├── right.mp4
    │   └── signals.npz
    ├── norm_stats.json
    ├── files.sha256.jsonl
    ├── metadata.json
    ├── audit.json
    ├── COMPLETE.json
    └── AUDIT_COMPLETE.json

低维轨迹格式

每条 raw/shards/*/episodes.jsonl 记录都包含稳定的 episode/task ID、split、任务类型、 中英文任务文本、声明式任务程序、合法和实际执行的子目标序列、确定性场景/专家随机种子、 初始状态指纹、环境重建参数、轨迹路径以及 replay 审计结果。

每个轨迹 NPZ 使用 T+1 个观测和 T 个动作：

数组	形状	含义
object_positions	[T+1, 9, 3]	9 个方块的位置
gripper_poses	[T+1, 2, 6]	左右夹爪的 xyz+rpy 位姿
suction	[T+1, 2]	夹爪开合状态
attached_objects	[T+1, 2]	仿真器中的抓取连接状态
actions	[T, 14]	左夹爪 7 维动作，随后是右夹爪 7 维动作
phase_ids	[T]	replay/审计阶段标签
segment_ids	[T]	语义片段编号

samples.parquet 提供可直接索引的策略窗口：

episode_id, task_id, split, segment_id,
frame_index, observation_frame_index,
proprioception[14], action_chunk[16,14], action_mask[16,14],
valid_steps, audit_phase_id

补齐后的动作必须结合 action_mask 使用，不能通过数值是否为零来推断有效位置。 segment_labels.jsonl 中的结构化事件和审计字段属于特权信息，不应作为部署策略输入。

RGB 格式

rgb_train_validation/episodes.jsonl 是公开索引。每条记录包含 episode_key、split、 任务文本、帧数/动作数、有效窗口数、signals 路径以及三路视频路径。

• 相机：global、left、right；
• 分辨率：256×192；帧率：20 FPS；
• 编码：无损 libx264rgb H.264，GOP 16；
• 左夹爪标记：亮紫色小球；右夹爪标记：亮橙色小球；
• 策略可见数据仅包括任务文本、视频、本体状态和动作。

下载与最小读取示例

仓库为私有数据集，需要使用具有访问权限的 Hugging Face 账号：

hf download zhb10086/umi-toybench-dual \
  --repo-type dataset \
  --local-dir data/umi-toybench-dual

只下载低维 benchmark：

hf download zhb10086/umi-toybench-dual \
  --repo-type dataset \
  --exclude 'rgb_train_validation/**' \
  --local-dir data/umi-toybench-dual

import json
from pathlib import Path
import pyarrow.parquet as pq

root = Path("data/umi-toybench-dual")
with (root / "rgb_train_validation/episodes.jsonl").open() as stream:
    episode = json.loads(next(stream))
print(episode["episode_key"], episode["split"], episode["videos"])

table = pq.read_table(root / "raw/shards/shard_00000/samples.parquet")
print(table.schema)
print(table.slice(0, 1).to_pylist()[0]["valid_steps"])

自包含使用说明

理解和消费本数据所需的定义均在数据集内：任务与 split 由 task_catalog.json 和 episode 记录确定；场景/渲染参数由 episode 记录和 protocol.json 确定；训练窗口、归一化、成功 判定、泄漏检查和内容哈希分别由 Parquet、统计文件、审计文件和 COMPLETE.json 给出。 使用者可以选择任意仿真器和策略框架，但必须保持上述数组约定和以下评测规则。

评测协议

1. 只能使用训练集和验证集训练模型、选择 checkpoint。
2. 查看测试结果前冻结 checkpoint、评测配置、episode manifest、随机种子和 rollout 预算。
3. 根据低维场景/任务记录重建测试 episode，并在线生成策略观测。
4. scorer truth、物体 ID、目标区域、阶段标签和连接状态不得进入策略输入。
5. 分 split 报告任务成功率、置信区间以及 ID 到各分布变化的性能差值。
6. 对组合、语言和视觉变化保留配对 episode 身份，执行 matched-scene 比较。

训练分布上的高 replay 成功率不能替代 held-out benchmark 结果。

审计与完整性

低维数据通过了全量 schema 检查、确定性物理 replay、64,100 个语义片段验证、 798,067 个 horizon 样本转换检查、1,200 个匹配组的 split/leakage 检查，以及 202 个 shard 的 checksum 验证。RGB 数据额外验证全部文件哈希、signals、视频时间轴、标记可见性 和归一化统计。

root COMPLETE.json                  a189c217c8207873a363f773feae13156a690c5af1299030b16c07061ffde907
raw physical content                6b27f81a32167ea33330f7ca9993c9d8eb9d0209f9791ded401472ccf3f8941c
RGB COMPLETE.json                   fd27416aa2984c6734c2525f02def00777dda017aa15c0432da95d57bc0565ec
raw/audit.json                      35c12b414a29fc35161bd2bcc6d64962a94b55f27e4e1a0a8b93cff70bd2e628
RGB AUDIT_COMPLETE.json             b09ef514ca2d345a0be89555a6f45f1492c7439b71232b3e4490a6335c5dbe0b

manifest 中历史构建环境的绝对路径不是下载后的必需路径；应以内容哈希识别本次发布。

限制与合理使用

• 数据全部来自 PyBullet 合成演示，尚未证明能够迁移到真实机器人。
• 物体、接触和连续夹爪运动均简化了真实硬件物理过程。
• 测试 RGB 需要在线生成，结果依赖固定的仿真和渲染约定。
• 低维记录为重建和审计保留了特权字段；将其输入模型会造成严重数据泄漏。
• 本数据集不是经过安全认证的机器人控制器。

许可证：Apache-2.0。使用者仍需遵守自行选择的模型和仿真器许可证，并在任何真实硬件 部署前单独完成安全验证。

---

English

umi-toybench-dual is a synthetic PyBullet benchmark for long-horizon, language-conditioned manipulation with two independently controlled grippers. The scene contains nine colored, size-varying blocks and three target regions. Each gripper is rendered as a pair of collision-free spheres, making jaw position and opening visible from three camera views.

The release supports controlled sanity checks, imitation learning, and ID/compositional/language/ semantic/visual generalization evaluation. The final public-facing data layout contains two repositories:

Dataset	Role
umi-toybench-dual	Complete low-dimensional benchmark and train/validation RGB
umi-toybench-dual-1000	The 1,000-episode experiment cohort, ACT training package, and timed latent-skill overlay

This repository contains generated benchmark data only. It does not mirror LIBERO, and it does not contain checkpoints, optimizer state, runtime logs, failed runs, or evaluation outputs.

Dataset summary

Package	Episodes	Actions / frames	Purpose
raw/	12,900	3,095,162 actions	Low-dimensional trajectories, task/scene definitions, segments, and horizon-16 samples
rgb_train_validation/	6,900	1,417,514 observation frames	Train/validation RGB, signals, and episode metadata

The raw package contains 64,100 verified semantic segments and 798,067 stride-4, horizon-16 sample records across 202 shards. The RGB package contains 34,500 episode files and approximately 6.57 GB of audited media/signals.

Splits and controlled shifts

Each task definition has 50 physical replicas. Train contains 132 task definitions; validation contains 6; each test family contains 24.

Split	Task definitions	Episodes	What changes
train	132	6,600	36 atomic + 84 seen ordered pairs + 12 multi-solution tasks
validation	6	300	Seen pair programs in new ID scenes
test_id	24	1,200	Seen programs in new ID scenes
test_composition	24	1,200	Held-out ordered conjunctions; both atomic marginals are seen
test_paraphrase	24	1,200	Seen programs with held-out language paraphrases
test_semantic	24	1,200	Unseen three-clause semantic macros composed from seen atomic clauses
test_visual	24	1,200	Seen programs with a held-out rendering domain

The five test families share matched structural scene groups where applicable. This permits paired comparisons in which composition, text, semantic abstraction, or rendering changes while the underlying scene structure is controlled.

Only train and validation have pre-rendered RGB. Test episodes intentionally remain low-dimensional: closed-loop evaluation must reconstruct the frozen scene and render observations online. This prevents accidental consumption of test images during training or model selection.

Directory layout

.
├── COMPLETE.json
├── metadata.json
├── RELEASE.json
├── raw/
│   ├── COMPLETE.json
│   ├── metadata.json
│   ├── protocol.json
│   ├── run_config.json
│   ├── task_catalog.json
│   ├── audit.json
│   └── shards/shard_*/
│       ├── episodes.jsonl
│       ├── segment_labels.jsonl
│       ├── samples.parquet
│       ├── trajectories/<split>/*.npz
│       ├── manifest.json
│       └── audit.json
└── rgb_train_validation/
    ├── episodes.jsonl
    ├── episodes/<prefix>/<episode_key>/
    │   ├── global.mp4
    │   ├── left.mp4
    │   ├── right.mp4
    │   └── signals.npz
    ├── norm_stats.json
    ├── files.sha256.jsonl
    ├── metadata.json
    ├── audit.json
    ├── COMPLETE.json
    └── AUDIT_COMPLETE.json

Low-dimensional schema

Each raw/shards/*/episodes.jsonl record binds:

• stable episode/task IDs, split, task kind, and English/Chinese task text;
• a declarative program and legal/realized subgoal sequences;
• deterministic scene and expert seeds plus an initial-state fingerprint;
• scene geometry/render-domain parameters used to reconstruct the simulator;
• trajectory path, action/frame/segment counts, and replay audit results.

Each trajectory NPZ uses T+1 observations and T actions:

Array	Shape	Meaning
object_positions	[T+1, 9, 3]	Nine block positions
gripper_poses	[T+1, 2, 6]	Left/right xyz+rpy poses
suction	[T+1, 2]	Gripper open/closed state
attached_objects	[T+1, 2]	Simulator attachment state
actions	[T, 14]	Left xyz+rpy+suction, followed by right xyz+rpy+suction
phase_ids	[T]	Replay/audit stage labels
segment_ids	[T]	Semantic segment membership

segment_labels.jsonl includes text supervision, valid next-subgoal alternatives, simulator event boundaries, and structured semantic/audit fields. The latter are privileged metadata and must not be fed to deployment policies.

samples.parquet contains the indexed policy windows:

episode_id, task_id, split, segment_id,
frame_index, observation_frame_index,
proprioception[14], action_chunk[16,14], action_mask[16,14],
valid_steps, audit_phase_id

Padded action values are accompanied by the explicit mask. Consumers should never infer validity from numeric zeros.

RGB episode schema

rgb_train_validation/episodes.jsonl is the public index. Each record provides episode_key, split, task, frame/action counts, valid window count, the signals path, and three video paths.

The visual contract is:

• cameras: global, left, right;
• resolution: 256x192;
• frame rate: 20 FPS;
• encoding: lossless libx264rgb H.264, GOP 16;
• random access: audited for Decord;
• left marker: bright purple sphere pair;
• right marker: bright orange sphere pair;
• marker position source: deployment-visible gripper proprioception.

The model-visible RGB package is limited to task text, videos, proprioception, and actions. It contains no private provenance or structured subgoal inputs.

Download

The repository is private and requires an authenticated Hugging Face account with access:

hf download zhb10086/umi-toybench-dual \
  --repo-type dataset \
  --local-dir data/umi-toybench-dual

For low-dimensional-only evaluation, exclude the media-heavy RGB directory:

hf download zhb10086/umi-toybench-dual \
  --repo-type dataset \
  --exclude 'rgb_train_validation/**' \
  --local-dir data/umi-toybench-dual

Minimal inspection

import json
from pathlib import Path

root = Path("data/umi-toybench-dual")
with (root / "rgb_train_validation/episodes.jsonl").open() as stream:
    episode = json.loads(next(stream))
print(episode["episode_key"], episode["split"], episode["videos"])

import pyarrow.parquet as pq

table = pq.read_table(
    "data/umi-toybench-dual/raw/shards/shard_00000/samples.parquet"
)
print(table.schema)
print(table.slice(0, 1).to_pylist()[0]["valid_steps"])

Self-contained use

Everything needed to inspect the demonstrations, build policy windows, reconstruct frozen test scenes, and score a rollout is described by files in this dataset:

Need	Dataset source of truth
Task and split membership	raw/task_catalog.json and episode records
Initial scene and renderer state	each episode record plus raw/protocol.json
Demonstration observations/actions	trajectory NPZ files
Semantic segments	segment_labels.jsonl
Fixed-horizon training examples	samples.parquet
Train/validation RGB	rgb_train_validation/episodes.jsonl and episodes/
Normalization	rgb_train_validation/norm_stats.json
Success and leakage checks	per-shard and package audit files
Content identity	the COMPLETE.json and SHA-256 manifests

A consumer does not need a separate project repository to interpret the data. Implementations may use any simulator and policy framework as long as they preserve the array conventions above and the evaluation rules below. The dataset deliberately excludes executable controllers and model checkpoints so benchmark data cannot silently select a preferred implementation.

Evaluation contract

1. Train and select checkpoints using train/validation only.
2. Freeze checkpoint, evaluation configuration, episode manifest, random seeds, and rollout budget before accessing test results.
3. Reconstruct test episodes from raw scene/program records and render observations online.
4. Keep scorer truth, object IDs, targets, phase IDs, and simulator attachment state outside the policy input.
5. Report task success per split with confidence intervals and the ID-to-shift gaps.
6. For matched composition/language/visual comparisons, retain paired episode identity.

The benchmark itself does not prescribe a single model architecture. ACT and latent-skill artifacts are documented together in the umi-toybench-dual-1000 card. A high training replay score must not be reported as held-out benchmark performance.

Audits and integrity

The raw package passed six full-dataset layers:

1. schema/static validation;
2. deterministic physics replay for every episode;
3. semantic-label validation for 64,100 segments;
4. horizon-chunk transformation for 798,067 samples;
5. split/leakage checks over 1,200 matched groups;
6. checksum validation across all 202 shards.

The RGB package additionally verifies every listed file checksum, all signal arrays against the source, all video timelines, marker visibility, and normalization statistics.

Release anchors:

root COMPLETE.json                  a189c217c8207873a363f773feae13156a690c5af1299030b16c07061ffde907
raw physical content                6b27f81a32167ea33330f7ca9993c9d8eb9d0209f9791ded401472ccf3f8941c
RGB COMPLETE.json                   fd27416aa2984c6734c2525f02def00777dda017aa15c0432da95d57bc0565ec
raw/audit.json                      35c12b414a29fc35161bd2bcc6d64962a94b55f27e4e1a0a8b93cff70bd2e628
RGB AUDIT_COMPLETE.json             b09ef514ca2d345a0be89555a6f45f1492c7439b71232b3e4490a6335c5dbe0b

Historical absolute *root values in manifests are build provenance, not required consumer paths. Use content hashes rather than those paths to identify a release.

Limitations and responsible use

• All data are synthetic PyBullet demonstrations; sim-to-real transfer is not established.
• Objects, grippers, contacts, and continuous jaw motion simplify physical hardware.
• Test RGB is generated online and therefore depends on the pinned simulator/render contract.
• Raw records intentionally retain privileged fields for reconstruction and auditing. A consumer can create severe leakage by passing them to a model.
• The dataset provides evaluation data and contracts, not a safety-certified controller.

License

Apache-2.0. Users are responsible for complying with any model or simulator licenses they choose and for performing separate safety validation before any physical deployment.