Dataset card template

Dataset card template#

Copy this file to the root of a demonstration release as README.md, fill in every <...> placeholder, and delete the <!-- how to fill --> blocks. Most values come from the batch’s metadata.json, or from meta/info.json and meta/source_metadata.json in a LeRobot v3 export. The pointers below name the field.

---
pretty_name: "BiGym 2.0: <task> (<robot>, <n> demos)"
license: cc-by-4.0
language:
  - en
task_categories:
  - robotics
tags:
  - LeRobot
  - bigym
  - humanoid
  - loco-manipulation
  - bi-manual-manipulation
  - imitation-learning
  - teleoperation
  - mujoco
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files: data/*/*.parquet
---

BiGym 2.0: <task_name>#

<one or two sentences: what the operator does in this task, and what counts as success.>

  • Homepage and code: swirl-uk/BiGym2

  • Format: LeRobot v3, lossless PNG image mode

At a glance#

Field

Value

Source

Episodes

<60>

info.json → total_episodes

Frames

<37705>

info.json → total_frames

Operator(s)

<name or anonymous id, and how episodes split between several>

collection.sessions[].operator

Collection dates

<YYYY-MM-DD> to <YYYY-MM-DD>

collection.sessions[].started_at/finished_at

Data license

CC BY 4.0

this card

Task and success criterion#

  • Success: the task predicate holds for <success_hold_seconds> consecutive seconds and the robot never falls (task.success_hold_seconds).

  • Collection hold: <collect_success_hold_seconds> s (task.collect_success_hold_seconds).

  • Reach tolerance (reach tasks only): <0.05> m (task.reach_tolerance).

  • Episode budget for training and evaluation: <episode_length> env steps, the length bigym.loco.make(task) sets.

  • Demo down-sample rate: <demo_down_sample_rate>.

Robot, controller and action space#

  • Robot: <robot_model>: <Unitree G1, 29 DoF, Dex1-1 two-finger grippers>

  • Lower body: <backend> policy, stepped inside the environment at the outer control rate. The agent sends base velocity and height commands and never actuates the legs directly.

  • Upper body: <upperbody_ik_backend> inverse kinematics from VR controller poses (metadata.json → action_semantics).

  • Outer action (action, dimension <outer_action_dim>): <4> base slots (<pelvis_x, pelvis_y, pelvis_z, pelvis_rz>), <14> arm joints and <2> gripper scalars. Copy outer_action_floating_dofs and the bounds in fullbody_layout.outer_action_low/high, and say whether the pitch slot is enabled.

  • Base velocity scale during collection: vx = <0.35>, vy = <0.25> m/s (action_semantics.base_velocity_scale). The yaw scale is the collector’s --base-wz-scale (default 0.5 rad/s) and is not recorded.

  • Normalization: the stored action is the raw outer action normalized with the collector’s action_stats.min/max (action_stats_mode = <collector_envelope>). Disclose any further training normalization separately.

Reset semantics#

Reproducing this data requires reproducing the reset (metadata.json → reset_semantics):

  • init_stance: <keyframe>

  • init_pelvis_z: <0.74> m

  • reset_warmup_steps: <200> (<4.0> s). Episodes settle before the operator engages.

  • demo_start: <post_warmup_vr_engage_state>

Substrate fingerprint#

Results are comparable when substrate_version matches. The main fields:

{
  "substrate_version": "<bigym2-mj381-v1>",
  "mujoco_version": "<3.8.1>",
  "robot_model": "<g1_dex1>",
  "lowerbody_backend": "<groot_wbc_g1>",
  "lowerbody_weights": <{"<file>": "<sha256>", ...}>,
  "g1_passive_base_tilt": <true>,
  "solver": <2>,
  "contact_signature": "<a3972624f79d3e98>"
}

The full fingerprint is in metadata.json → substrate_fingerprint.

Package versions at collection time (metadata.json → package_versions): bigym <1.0.0>, mujoco <3.8.1>, mink <1.2.0>, numpy <2.2.6>, python <3.12.3>.

Collection#

  • Method: VR teleoperation with bigym-collect, operator wearing <headset>.

  • Attempts and kept: <attempts> attempts, <kept> kept (<xx>%), from collection.attempts and collection.kept.

  • Discarded attempts: <not released | released under data/discarded/>

Dataset structure#

Standard LeRobot v3 layout, plus these files:

<dataset>/
  meta/
    episode_init_states.json   # engage snapshots (BiGym sidecar)
    alignment.json             # action alignment version (BiGym sidecar)
    source_metadata.json       # the collector's metadata.json
  metadata.json                # root copy of source_metadata.json

Per-frame fields#

Field

dtype

shape

Notes

observation.state

float32

(<50>,)

low-dim observation, slices in substrate_fingerprint.low_dim_component_slices

observation.images.head

image

(3, <84>, <84>)

head camera, RGB, lossless PNG

observation.images.left_wrist

image

(3, <84>, <84>)

left wrist camera

observation.images.right_wrist

image

(3, <84>, <84>)

right wrist camera

action

float32

(<20>,)

normalized outer action, the training target

raw_outer_action

float32

(<20>,)

outer action before normalization

expanded_action

float32

(<35>,)

outer action expanded to the full joint set

lowerbody_action

float32

(<15>,)

lower-body policy output

leg_joint_targets

float32

(<15>,)

leg and waist position targets sent to the actuators

torso_target

float32

(1,)

torso yaw target

lowerbody_command

float32

(3,)

base velocity command (vx, vy, wz)

height_command

float32

(1,)

pelvis height command, m

full_qpos

float64

(<46>,)

full MuJoCo qpos

full_qvel

float64

(<45>,)

full MuJoCo qvel

reward

float32

(1,)

task reward

event_progress

float32

(1,)

per-task sub-goal progress, all NaN on tasks without events

discount

float32

(1,)

wall_clock

float64

(1,)

seconds since collector start, a diagnostic never used for training or replay

Action alignment#

Frame k pairs its observation with the transition executed from it (meta/alignment.json, version 2). The final frame repeats the last transition.

Engage snapshots#

meta/episode_init_states.json stores each episode’s engage snapshot: the seed and the simulator and controller state needed to replay it bit-exactly. Consumers that ignore it still get valid trajectories.

Intended use and limitations#

  • Task coverage: <list the tasks actually present in this release>.

  • Scale: <n> episodes from <one operator>. Expect operator-specific strategies and limited coverage of the task’s state space.

License#

  • Data (this dataset): CC BY 4.0. Attribute BiGym 2.0 and cite the entries below.

  • Code: the BiGym 2.0 repository is Apache-2.0. Third-party components redistributed with the code (robot models, lower-body policy weights, 3D props) have their own terms. See THIRD_PARTY_NOTICES.md in the repository.

  • Scene assets visible in the images: the rendered observations contain 3D props under CC0 and CC BY 4.0. Their authors are credited in bigym/envs/xmls/3D_MODELS_ATTRIBUTION.md, and that attribution carries over to anyone redistributing these images.

Citation#

Cite both the BiGym 2.0 paper and the original BiGym paper:

@article{zhang2026bigym2,
  title   = {BiGym 2.0: Benchmarking Learned and Agent-Developed Policies for Humanoid Household Manipulation},
  author  = {Zhang, Zexi and Zhu, Zecheng and Chen, Zidong and Tuya, Zulkhuu and James, Stephen},
  journal = {arXiv preprint arXiv:2610.07594},
  year    = {2026}
}

@article{chernyadev2024bigym,
  title   = {{BiGym}: A Demo-Driven Mobile Bi-Manual Manipulation Benchmark},
  author  = {Chernyadev, Nikita and Backshall, Nicholas and Ma, Xiao and Lu, Yunfan and Seo, Younggyo and James, Stephen},
  journal = {arXiv preprint arXiv:2407.07788},
  year    = {2024}
}

If the release uses the groot_wbc_g1 lower-body backend, also credit NVIDIA’s GR00T Whole-Body Control policy. THIRD_PARTY_NOTICES.md §1 has the terms its weights are distributed under.

Contact#

<maintainer name / issue tracker URL>