MCAP
MCAP is an open, modular container file format for recording multiple channels of timestamped, pre-serialized data, commonly used for robotics and multimodal logs.
For AI leaders, multimodal and robotics teams, data operations, evaluation teams, and technical buyers
Definition: MCAP is an open, modular container file format for recording multiple channels of timestamped, pre-serialized data, commonly used for robotics and multimodal logs.
Category: Physical AI and robotics
Full Definition
An MCAP file can store messages from many channels while retaining schemas, channel metadata, log time, publish time, optional attachments, metadata records, summaries, indexes, and compression. It is serialization-agnostic: payloads may use ROS 1, ROS 2 CDR, Protobuf, JSON Schema, FlatBuffers, or other registered encodings. The format is designed for efficient append-oriented writing, inspection, seeking, and recovery under varied resource constraints.
MCAP is a container, not a complete robotics dataset schema. It does not by itself define what a topic means, which clock is authoritative, whether sensors are calibrated, whether an episode is complete, or whether data may be used for training. Those requirements belong in message definitions, metadata, manifests, calibration artifacts, and program documentation.
How It Works in Practice
A recorder writes channel and schema definitions followed by timestamped messages. Implementations can chunk, compress, index, and summarize data for performance. Attachments can carry files such as calibration or images, while metadata records can describe the log. The official CLI can inspect, validate, filter, merge, sort, convert, compress, recover, and report on files.
For AI data delivery, pair MCAP with an episode manifest that states task, robot, environment, source class, required channels, clock model, calibration IDs, start/end, outcome, interventions, rights, quality state, and hashes. Validate that files open, schemas resolve, timestamps are monotonic where expected, required channels overlap, indexes are usable, and message counts and rates match the manifest.
Why It Matters for AI Data
MCAP can reduce custom-format fragmentation and preserve heterogeneous streams in one portable log. It is useful for robotics collection, replay, annotation, visualization, and exchange. Buyers should still specify semantic and quality requirements; accepting “MCAP delivery” without a channel contract and loader validation leaves critical ambiguity.
What a Production Record May Contain
| Field or artifact | Purpose |
|---|---|
| File identity | URI, content hash, size, creation, writer/library version, and compression. |
| Channels and schemas | Topic/channel, message encoding, schema ID, semantic owner, unit, and frame. |
| Timing | Log/publish time, clock source, start/end, ordering, skew, and correction metadata. |
| Episode context | Task, robot, site, calibration, required channels, outcome, and intervention. |
| QA and governance | Validation report, privacy/rights class, retention, release, and limitations. |
Quality and Governance Risks
- A syntactically valid file can contain missing channels, wrong units, unknown frames, or unsynchronized clocks.
- Large logs can expose people, locations, proprietary environments, credentials, or debugging content.
- Channel names and schemas can drift across firmware or collection sites.
- Attachments and metadata may be omitted even when required to interpret the messages.
- Compression and chunk settings affect write durability, random access, and processing cost.
- Converting from another format can lose schemas, timestamps, metadata, or original ordering unless validated.
Practical Example
A manipulation episode is delivered as one MCAP file containing RGB, depth, joint state, end-effector pose, force/torque, action commands, controller events, and safety events. A sidecar manifest pins every schema and unit, declares the master clock and calibration records, lists expected rates and maximum gaps, identifies terminal outcome, and includes file and channel hashes. Automated acceptance verifies both the MCAP structure and the episode contract.
Related Terms
ROS Bag · Sensor Fusion · VLA · LiDAR Annotation
Key Takeaway
MCAP standardizes how heterogeneous timestamped messages are contained; it does not replace task, schema, synchronization, calibration, quality, or governance specifications.