Training Formats

Training-ready in the format your stack expects.

LeRobot v2/v2.1, GR00T-LeRobot with modality maps, ALOHA-style HDF5, MCAP for ROS 2, RLDS for OXE pipelines, robomimic for clean baselines. Required fields, optional fields, and what SignIQ exports are documented per format below.

Compatibility at a glance

Pick the format that matches your training stack. SignIQ exports the four green-checked rows directly; RLDS conversion is available via the LeRobot mirror.

FormatBest forVideoState / actionUsed by
LeRobotModern HF/PyTorch roboticsyesyesLeRobot, openpi, GR00T adapters
GR00T-LeRobotNVIDIA GR00T fine-tuningyesyesIsaac GR00T
HDF5ACT, Diffusion Policy, robomimicoptionalyesACT, robomimic, ALOHA
MCAPROS-native logsyesyesRobotics infra / replay
RLDSOXE/RT-X/Octo/JAXyesyesJAX/TF robot learning

LeRobot (v2 / v2.1)

SignIQ exports

Hugging Face's PyTorch-native robot dataset format. Parquet for low-dim, MP4 for video, JSONL/JSON metadata. Modern default for VLA work.

File structure

meta/info.json
meta/episodes.jsonl
meta/tasks.jsonl
meta/stats.json
data/chunk-XXX/episode_XXX.parquet
videos/chunk-XXX/<camera>/episode_XXX.mp4

Required fields

  • observation.images.<camera>
  • observation.state
  • action
  • timestamp
  • frame_index
  • episode_index

Optional fields

task (language instruction) · next.reward · next.done · annotation.<custom>

Action representation
Per-dataset; documented in meta/info.json features
Best for
openpi / GR00T / OpenVLA / LeRobot fine-tuning and PyTorch pipelines
Used by
LeRobot, openpi, GR00T (with modality.json), many OXE mirrors
SignIQ export
Yes — shipped natively

GR00T-LeRobot

SignIQ exports

LeRobot v2 with an additional meta/modality.json that maps state, action, and video keys to GR00T's policy expectations. Required for Isaac-GR00T fine-tuning.

File structure

All LeRobot v2 files
meta/modality.json (state.<part>, action.<part>, video.<view>)
meta/embodiment.json

Required fields

  • modality.state.<body_part>
  • modality.action.<body_part>
  • modality.video.<camera>
  • embodiment tag

Optional fields

language · reward · subtask spans

Action representation
Per body part; GR00T handles chunking and tokenization internally
Best for
GR00T N1 / N1.7 fine-tuning on humanoid and bimanual platforms
Used by
NVIDIA GR00T, Isaac Lab post-training
SignIQ export
Yes — shipped natively

HDF5 (ALOHA / robomimic / ACT)

SignIQ exports

Hierarchical binary format with embedded multi-camera image stacks and joint trajectories. The native format for ACT, robomimic, and many bimanual recipes.

File structure

/observations/images/<camera>
/observations/qpos
/observations/qvel
/action
/episode_metadata

Required fields

  • observations.qpos
  • action
  • episode metadata

Optional fields

observations.images.<camera> · observations.qvel · observations.gripper · language_instruction

Action representation
Joint targets or action chunks; embedded shape/length per dataset
Best for
ACT, Diffusion Policy, robomimic baselines, ALOHA tooling
Used by
ACT, robomimic, Diffusion Policy, ALOHA scripts
SignIQ export
Yes — shipped natively

MCAP / ROS 2

SignIQ exports

ROS-native log container. Best for replay, debugging, and end-to-end systems work where the full message bus matters.

File structure

<bag>.mcap (channel-keyed messages)
<bag>.metadata.yaml

Required fields

  • /tf and /tf_static
  • /joint_states
  • /<camera>/image_raw

Optional fields

/wrench · /audio · /diagnostics · /<custom>

Action representation
Joint commands or twist/wrench messages on a control topic
Best for
Robotics infra teams, replay/visualization, ROS-first stacks
Used by
Foxglove, rosbag2, ROS 2 native pipelines
SignIQ export
Yes — shipped natively

RLDS / TFRecord

RLDS is the canonical format behind Open X-Embodiment. JAX/TF pipelines (Octo, OpenVLA) consume it directly.

File structure

*.tfrecord shards
features.json
dataset_info.json

Required fields

  • steps[].observation
  • steps[].action
  • steps[].is_terminal
  • steps[].language_instruction (optional)

Optional fields

episode_metadata · discount · reward

Action representation
Most VLA work uses 7D EEF delta + gripper
Best for
OXE/RT-X, OpenVLA, Octo, JAX/TF pretraining
Used by
Open X-Embodiment, OpenVLA, Octo, RT-X
SignIQ export
On request

robomimic

SignIQ exports

Curated HDF5 layout with proficient-human, multi-human, and machine-generated splits. Strong baselines and clean evaluation.

File structure

data/demo_<i>/obs/<key>
data/demo_<i>/actions
data/demo_<i>/dones
mask/train, mask/valid

Required fields

  • obs.<key>
  • actions
  • dones

Optional fields

rewards · next_obs · states (sim)

Action representation
Joint deltas or absolute targets per dataset config
Best for
Offline imitation learning, BC/diffusion baselines, ablations
Used by
robomimic, robosuite, diffusion policy baselines
SignIQ export
Yes — shipped natively

Need a format we don't list?

We can also export to robomimic-flavored HDF5, custom Parquet schemas, or embodiment-specific GR00T modality maps. Tell us your stack and we'll match it.