Task JSON reference
Read one task first, then use this reference to interpret its fields.
This reference covers every field observed in 61 inspected JSON examples across all 59 task-bearing directories, checked against the project appendix. It is not an exhaustive schema validation of all 773,716 tasks. An observed field is not automatically mandatory for every record.
All Data/... paths resolve against the extracted dataset root. Empty strings, empty lists, and the literal "none" are distinct recorded values. Missing fields should be handled explicitly.
Field dictionary
| Field | Observed type | Meaning and use | Example |
|---|---|---|---|
metadata |
object | Task identity and source-reference information. These fields are not automatically model inputs. | {…} · JSON |
metadata.task_id |
string | Task identifier. Preserve the full string; do not infer paths solely from its naming pattern. | "L1_audio_to_motion_generation_audio_--04DHOI32k_00009_38_121" · JSON |
metadata.task_level |
string | Benchmark level: Level1, Level2, or Level3. | "Level1" · JSON |
metadata.task_family |
string | Human-readable task family. The task map links this label to its exact directory spelling. | "Audio-to-Motion Generation" · JSON |
metadata.task_type |
string | Task-specific variant or constraint label (for example fast, scale_up, legs, or 2x). An empty string means no label is recorded here. Read the instruction for the actual requested behavior. | "" · JSON |
metadata.original_video |
string | Dataset-root-relative source/reference video path. May refer to human or skeleton footage. It is not necessarily an allowed conditioning input. | "Data/Shared/Video/Human/00065/--04DHOI32k_00009_38_121.mp4" · JSON |
metadata.original_motion |
string | Dataset-root-relative source/reference motion path: usually a CSV directory, but a PKL file for Order/Times. | "Data/Shared/Motion/00065/--04DHOI32k_00009_38_121" · JSON |
metadata.duration |
number | Recorded task/source duration in seconds. Do not infer CSV row count or modified output duration from this alone; motion packages can include a hold period. | 2.803 · JSON |
input |
object | Conditioning information supplied for this task. | {…} · JSON |
input.prompts |
object | Instruction strings. Use together with the modality content. | {…} · JSON |
input.prompts.general_instruction |
string | Task-level instruction describing the expected behavior. Preserved verbatim, including any output-format wording; no tokenizer or model interface is supplied here. | "You are an advanced 3D motion generation model designed to synthesize p… · JSON |
input.prompts.modifier_instruction |
string | Additional behavioral constraint, when present. Empty does not imply no constraint: it may be contained in text or spoken audio. | "" · JSON |
input.modalities |
object | Level 1/2 conditioning fields. In the inspected Level 3 examples this is empty; use input.interleave.sequence instead. | {…} · JSON |
input.modalities.text |
string | Inline text condition. Empty string means no text content in this field; audio may carry the instruction. | "" · JSON |
input.modalities.audios |
array | Ordered list of dataset-root-relative audio file paths. Empty list means this field supplies no audio. | […] · JSON |
input.modalities.videos_processed |
array | Ordered list of conditioning-video paths. Order/video uses two clips whose list order is referenced by the instruction. Do not replace these with metadata.original_video. | […] · JSON |
input.modalities.spatial_coordinates |
array | List of conditioning-file paths, not inline coordinate arrays. Rotation-to-Pose examples point to a PKL motion-conditioning pack. | […] · JSON |
input.modalities.input_images |
array | List of image conditions. Key-frame examples use objects containing path and time; Trajectory examples use path strings. Support both forms. | […] · JSON |
ground_truth |
object | Task reference/target information. Its meaning depends on the task family; it is not a conditioning input. | {…} · JSON |
ground_truth.motion_parameters |
string | Motion directory or file referenced by this task. Temporal-completion examples point to the missing segment; Amplitude, Speed, Direction, and Body Restrain point to the unmodified reference motion. | "Data/Shared/Motion/00065/--04DHOI32k_00009_38_121" · JSON |
ground_truth.rendered_video |
string | Video path recorded with the reference, or the literal string "none" when absent. Do not treat "none" as a filename. A video here is not necessarily a rendered constraint-satisfying output. | "none" · JSON |
input.modalities.input_images[] |
object | Image-condition object; see path and time below. | {…} · JSON |
input.modalities.input_images[].path |
string | Key-frame image path relative to the dataset root. | "Data/Level1/Image/HumanKeyframes/--3wjNOccLY_00002_0_63_joint0.jpg" · JSON |
input.modalities.input_images[].time |
number | Key-frame timestamp in seconds on the source motion timeline. | 1.2 · JSON |
ground_truth.time_window |
array | Two-element [start, end] interval in seconds on the original timeline, recorded for temporal-completion targets. The referenced motion directory is already segmented. Do not crop it a second time. Endpoint-to-frame details must follow the package and evaluation code. | […] · JSON |
ground_truth.trajectory_points |
object | Trajectory-task reference waypoints in two coordinate representations. | {…} · JSON |
ground_truth.trajectory_points.canonical_xz_meters |
object | Named two-number [x, z] waypoints in the canonical horizontal plane, in meters. | {…} · JSON |
ground_truth.trajectory_points.canonical_xz_meters.start |
array | Two-number start waypoint, using the coordinate representation of its parent. These names do not by themselves specify timestamps or an interpolation rule. | […] · JSON |
ground_truth.trajectory_points.canonical_xz_meters.point_1_4 |
array | Two-number quarter-position waypoint, using the coordinate representation of its parent. These names do not by themselves specify timestamps or an interpolation rule. | […] · JSON |
ground_truth.trajectory_points.canonical_xz_meters.point_1_2 |
array | Two-number midpoint waypoint, using the coordinate representation of its parent. These names do not by themselves specify timestamps or an interpolation rule. | […] · JSON |
ground_truth.trajectory_points.canonical_xz_meters.point_3_4 |
array | Two-number three-quarter-position waypoint, using the coordinate representation of its parent. These names do not by themselves specify timestamps or an interpolation rule. | […] · JSON |
ground_truth.trajectory_points.canonical_xz_meters.end |
array | Two-number end waypoint, using the coordinate representation of its parent. These names do not by themselves specify timestamps or an interpolation rule. | […] · JSON |
ground_truth.trajectory_points.image_pixels |
object | Corresponding two-number image-coordinate waypoints in pixels. Keep the paired image; do not interpret these as meters. | {…} · JSON |
ground_truth.trajectory_points.image_pixels.start |
array | Two-number start waypoint, using the coordinate representation of its parent. These names do not by themselves specify timestamps or an interpolation rule. | […] · JSON |
ground_truth.trajectory_points.image_pixels.point_1_4 |
array | Two-number quarter-position waypoint, using the coordinate representation of its parent. These names do not by themselves specify timestamps or an interpolation rule. | […] · JSON |
ground_truth.trajectory_points.image_pixels.point_1_2 |
array | Two-number midpoint waypoint, using the coordinate representation of its parent. These names do not by themselves specify timestamps or an interpolation rule. | […] · JSON |
ground_truth.trajectory_points.image_pixels.point_3_4 |
array | Two-number three-quarter-position waypoint, using the coordinate representation of its parent. These names do not by themselves specify timestamps or an interpolation rule. | […] · JSON |
ground_truth.trajectory_points.image_pixels.end |
array | Two-number end waypoint, using the coordinate representation of its parent. These names do not by themselves specify timestamps or an interpolation rule. | […] · JSON |
input.interleave |
object | Level 3 ordered multimodal condition and its recorded selection identifiers. | {…} · JSON |
input.interleave.seed |
integer | Recorded integer seed. Preserve as an integer; observed values can exceed signed 64-bit or JavaScript exact-integer range. Not a timing value or input asset. | 16934628329455981502 · JSON |
input.interleave.seed_source |
string | String identifying the recorded seed source. Preserved for traceability; not a path to download. | "--6s5bu1NRU_00008_0_463 / L3_INTERLEAVE / v1" · JSON |
input.interleave.sequence |
array | Chronologically ordered list of modality components. Retain order and timestamps when reading. | […] · JSON |
input.interleave.sequence[] |
object | One typed condition component. Fields depend on type; not every component has every field below. | {…} · JSON |
input.interleave.sequence[].order |
integer | Zero-based position in the interleaved component sequence. | 0 · JSON |
input.interleave.sequence[].type |
string | Component modality: text, audio, video, or image. | "video" · JSON |
input.interleave.sequence[].segment_index |
integer | Zero-based motion-segment index for text/audio/video. Different from order because images may be inserted between segments. | 0 · JSON |
input.interleave.sequence[].start_time |
number | Segment start time in seconds on the original timeline. | 0.0 · JSON |
input.interleave.sequence[].end_time |
number | Segment end time in seconds on the original timeline. This does not specify a CSV slicing convention by itself. | 3.5 · JSON |
input.interleave.sequence[].video_path |
string | For video components: dataset-root-relative cropped input-video path. | "Data/Level3/Video/--6s5bu1NRU_00008_0_463_seg00.mp4" · JSON |
input.interleave.sequence[].boundary_after_segment |
integer | For image components: zero-based index of the segment preceding this boundary. | 0 · JSON |
input.interleave.sequence[].timestamp |
number | For image components: boundary timestamp in seconds on the original timeline. | 3.5 · JSON |
input.interleave.sequence[].image_path |
string | For image components: dataset-root-relative boundary-image path. | "Data/Level3/Image/--6s5bu1NRU_00008_0_463_boundary00.jpg" · JSON |
input.interleave.sequence[].audio_path |
string | For audio components: dataset-root-relative spoken segment-description path. | "Data/Level3/Audio/--6s5bu1NRU_00008_0_463_seg01.mp3" · JSON |
input.interleave.sequence[].text |
string | For text components: inline motion description. | "The person bends slightly forward to pick up a pneumatic impact wrench,… · JSON |
Task-specific interpretation
- Temporal completion: use the target directory and its original-timeline
time_window; metadata can still describe the full source clip. - Key frames: keep each image paired with its time. Trajectory image inputs have a different representation.
- Order: retain both input clips in list order. The full reference video is not a substitute for the two conditioning clips.
- Times: the reference PKL encodes repetitions; the recorded video may show the single source action. Do not label that video as a repeated result.
- Amplitude / Speed / Direction / Body Restrain: the reference motion is unmodified. The instruction specifies the requested change; the reference video is not a model result.
- Trajectory: use the condition image and recorded waypoints together; do not infer a coordinate transform solely from their values.
- Level 3: read
input.interleave.sequenceeven wheninput.modalitiesis empty. Keep the source timeline, segment indices, and image boundaries.
Motion files
| File | Contents |
|---|---|
joint_pos.csv |
29 robot joint-position columns |
joint_vel.csv |
Corresponding joint velocities |
body_pos.csv |
Root position in world coordinates |
body_quat.csv |
Root orientation quaternion |
body_lin_vel.csv |
Root linear velocity |
body_ang_vel.csv |
Root angular velocity |
The G1 packages described in the appendix use six synchronized CSV files at 50 Hz. Read each CSV header and the package metadata for its conventions. The inspected full-motion package includes a one-second initial hold, so task duration alone does not determine row count. Do not infer named joint order or quaternion output order from array length. Order/Times instead reference AMASS/BABEL-linked PKL packages; Rotation-to-Pose PKL conditions are another format. These are not interchangeable.
Some released prompt wording differs across task variants (including output-format wording). Examples here preserve those strings rather than silently rewriting task definitions. Model integration and official evaluation commands will be supplied with the benchmark code.