Yinpei/RoboMME_real_robot_data

RoboMME Real-Robot Data

This dataset contains real-robot camera observations, state vectors, action chunks, and language annotations for block selection, hidden-object memory, motion-pattern imitation, and object-counting tasks.

The repository was previously named Yinpei/HistoryVLARealRobotData. The archive path and internal directory names retain their original names.

Dataset at a glance

Property	Value
Archive	historyvla/real_robot_0101.tar.gz
Compressed size	34,777,720,984 bytes, approximately 34.8 GB
Extracted file contents	46,369,654,016 bytes, approximately 46.4 GB
Timestep records	78,400
Episodes	353
Distinct task prompts	10
Stored timesteps per episode	59–907
Camera observations	Three RGB images per timestep, each 256 × 256
State shape	(8,)
Action-chunk shape	(20, 8)

The counts and indexing properties in this README were checked across the complete archive. The payload was inspected at repository revision 5f98d6a41966046fffbe5c930a05cf3e69c05f0b.

Archive SHA-256:

c26c891210bec95ad79d9ea12e812d7d5dc29cf681e379b68d37e453f88d9397

Files and episode organization

historyvla/real_robot_0101.tar.gz
└── real_robot_0101/
    ├── meta/
    │   └── stats.json          # {"total_samples": 78400}
    └── data/
        ├── 0.pkl
        ├── 1.pkl
        ├── ...
        └── 78399.pkl

There are 78,401 files inside the archive: 78,400 pickle files and one metadata JSON file. Each .pkl is a dictionary for one timestep, not one complete episode. Images are NumPy arrays inside the pickles; the archive contains no separate image or video files.

The filename stem is a global sample ID. Episode membership and timing are stored in epis_idx and step_idx. Archive members appear in shuffled order. In this release, sorting filenames numerically produces consecutive episode segments and chronological steps within each episode. Lexicographic sorting would put 10.pkl before 2.pkl and should not be used.

Understanding the indices

Field	Meaning observed in this release
Filename, e.g. 297.pkl	Global sample ID, from 0 to 78,399
epis_idx	Episode ID, from 0 to 352
step_idx	Timestep index within the episode's original step numbering
exec_start_idx	Constant within an episode; equals that episode's first stored step_idx

exec_start_idx is not an index into the global list of pickle files. It uses the same episode-local numbering as step_idx.

File	epis_idx	step_idx	exec_start_idx	Relative stored step
0.pkl	0	157	157	0
1.pkl	0	158	157	1
2.pkl	0	159	157	2
295.pkl	0	452	157	295
296.pkl	0	453	157	296
297.pkl	1	111	111	0
298.pkl	1	112	111	1
476.pkl	1	290	111	179
477.pkl	2	170	170	0
478.pkl	2	171	170	1
57548.pkl	234	0	0	0

For all 353 episodes:

relative_step = step_idx - exec_start_idx

The filename ID continues increasing across episode boundaries. epis_idx increases by one at each boundary; step_idx starts at the new episode's exec_start_idx, which is not necessarily zero. There are no gaps in the stored step sequence within any episode.

History availability: 303 episodes start above step zero. No stored record has step_idx < exec_start_idx. For example, episode 0 contains steps 157–453, but steps 0–156 are not present. Prompts referring to an earlier video or previously picked objects do not imply that the preceding visual context is included in this archive.

Timestep schema

Every record has these 15 fields:

Field	Type and shape	Contents
left_shoulder_image	uint8 (256, 256, 3)	Left-shoulder RGB view
right_shoulder_image	uint8 (256, 256, 3)	Right-shoulder RGB view
wrist_image	uint8 (256, 256, 3)	Wrist RGB view
state	float32 (8,)	Robot state vector
actions	float32 (20, 8)	Twenty 8-value action vectors
is_demo	bool (1,)	Stored demonstration flag; false in every record
exec_start_idx	int32 (1,)	Start index described above
step_idx	int32 (1,)	Episode-local timestep index
epis_idx	int32 (1,)	Episode identifier
prompt	str	Task instruction
simple_subgoal	str	Subgoal annotation
grounded_subgoal	str	Grounded subgoal annotation; may include coordinate pairs
simple_subgoal_online	str	Online subgoal annotation
grounded_subgoal_online	str	Online grounded subgoal annotation
eef_vel_history	float32 (1, 4)	Zero in every record in this release

Identifiers and flags are one-element arrays, rather than Python scalars. Use .item() to read their scalar values. An is_demo=False value alone does not establish whether a human or policy generated the episode.

For 0.pkl, the task is:

pick up all the blocks that have been picked before.

All four subgoal fields in this sample contain pick up the block at <89, 165>. The coordinate convention and camera association of such coordinate pairs are not specified in the release.

State and action interpretation

The stored shapes and values are directly verified. The following semantic interpretation is inferred from numerical comparisons and camera observations; it is not a documented controller specification:

state[0:7]       likely measured arm joint positions
state[7]         consistent with normalized gripper closedness
actions[h, 0:7]  likely absolute arm joint-position targets
actions[h, 7]    binary gripper command
h = 0, ..., 19

The first seven action values closely track subsequent state values and have position-like offsets. Their aggregate RMSE against the current state is 0.03723 raw units, compared with 1.21410 against the next state difference. This supports an absolute-target interpretation rather than treating them as per-step deltas. Exact joint names, ordering, physical units, normalization, and controller API are not specified; radians must not be assumed to be a confirmed unit.

The last action dimension contains exactly 0 and 1. In the visually reviewed episodes (0 and 234), 0 corresponds to opening and 1 to closing. The last state dimension ranges from approximately 0.001596 to 0.998350 and is consistent with open-to-closed gripper state. It should not be interpreted as a documented width in meters.

Action chunks and episode ends

The 20 action rows behave predominantly as overlapping future-action windows: actions[t, 1] exactly matches actions[t + 1, 0] in 77,717 of 78,047 valid within-episode comparisons (99.58%). At offset 19, exact agreement is 97.81%.

There are exceptions, including away from episode ends. Keep the supplied chunks rather than silently rebuilding them from the first action of each record. Only 246 of 353 final records repeat their first action across all 20 rows, so uniform terminal padding must not be assumed. The precise chunk-construction procedure is not specified in the release.

Subgoal versions

Offline and online subgoal strings differ in 4,760 records (6.07%), for both the simple and grounded fields. They are separate annotations, not interchangeable copies. In episodes 0 and 234, online transitions occur ten stored steps before the corresponding offline transitions. Preserve both versions and select the intended annotation source explicitly.

Tasks

Task prompt	Episodes
pick up all the blocks that have been picked before.	89
pick up the block that has been picked before.	11
put 1 fruit from the basket to the bin, and press the button to stop.	13
put 2 fruits from the basket to the bin, and press the button to stop.	13
put 3 fruits from the basket to the bin, and press the button to stop.	13
put 4 fruits from the basket to the bin, and press the button to stop.	11
watch the video, close the gripper and replicate the same pattern.	103
watch the video, pick up the cup that hides the green block.	33
watch the video, pick up the cup that hides the red block.	33
watch the video, pick up the cup that hides the yellow block.	34

These total 100 block episodes, 100 cup/hidden-block episodes, 103 motion-pattern episodes, and 50 fruit-counting/button episodes.

Download and extract

If the repository requires access approval, obtain access on Hugging Face and authenticate using hf auth login before downloading.

pip install huggingface_hub numpy

from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="Yinpei/RoboMME_real_robot_data",
    repo_type="dataset",
    filename="historyvla/real_robot_0101.tar.gz",
    revision="5f98d6a41966046fffbe5c930a05cf3e69c05f0b",
    local_dir="downloads",
)
print(archive)

mkdir -p extracted
tar -xzf downloads/historyvla/real_robot_0101.tar.gz -C extracted

Keeping both the archive and extracted files requires approximately 81.2 GB, plus filesystem overhead and any download-cache overhead.

Load a timestep

Python pickle can execute code during loading. Use this example only with files from a source you trust. NumPy must be installed to deserialize the arrays.

import pickle
from pathlib import Path

data_dir = Path("extracted/real_robot_0101/data")
with (data_dir / "0.pkl").open("rb") as stream:
    sample = pickle.load(stream)

episode_id = sample["epis_idx"].item()       # 0
step = sample["step_idx"].item()             # 157
execution_start = sample["exec_start_idx"].item()  # 157
relative_step = step - execution_start      # 0

print(sample["prompt"])
print(sample["left_shoulder_image"].shape)  # (256, 256, 3)
print(sample["state"].shape)                # (8,)
print(sample["actions"].shape)              # (20, 8)

Group timesteps into episodes

This builds a lightweight index of file paths. It reads each pickle once, but does not retain all image arrays in memory. A complete scan reads approximately 46.4 GB of extracted file contents.

import pickle
from collections import defaultdict
from pathlib import Path

data_dir = Path("extracted/real_robot_0101/data")
episodes = defaultdict(list)

for path in sorted(data_dir.glob("*.pkl"), key=lambda p: int(p.stem)):
    with path.open("rb") as stream:
        sample = pickle.load(stream)
    episode_id = sample["epis_idx"].item()
    step = sample["step_idx"].item()
    episodes[episode_id].append((step, path))

for records in episodes.values():
    records.sort(key=lambda item: item[0])

print(len(episodes))        # 353
print(len(episodes[0]))     # 297
print(episodes[0][0][0])    # 157
print(episodes[0][-1][0])   # 453

For training/evaluation splits, split by episode so overlapping action windows from the same trajectory do not appear on both sides of the split.

Timing and evaluation limitations

• The release does not specify native frame rate, control frequency, timestamps, camera calibration, or a complete robot/controller contract.
• Videos made from the camera arrays require a chosen playback rate. That rate should be reported as a visualization setting, not as measured physical timing.
• No reward, success flag, or termination-reason field is present in the timestep dictionaries. A rendered camera video does not independently establish task success or constitute execution of the stored actions.
• Missing pre-execution history limits evaluation of instructions that depend on previously observed actions or hidden objects.

License

The dataset's existing license metadata declares Apache-2.0. This README preserves that declaration.