Robotics Data Tools and Platforms: How Teams Log, Search, Label, and Evaluate Robot Data

Nov 17, 2026
•
13 min read
Robotics data tools, also called robotics data platforms or physical AI data platforms, are the software robotics and physical AI teams use to record robot data, replay it, find the episodes that matter, and turn them into labeled training data and model evaluations. No single tool covers every stage well, so most teams pair two kinds. Tools built around the recording, like Foxglove, Rerun, and Roboto, get data off robots and help you debug it. Tools built around the dataset and the model, like Encord, Lightly, and Voxel51, turn recordings into training data and show where models fail. The groups overlap. Foxglove, Rerun, and Roboto also search and curate recordings, and Rerun and Roboto export LeRobot datasets.
This guide compares all six by job. It covers software, not companies that collect or label data for you. We make Voxel51, so every claim about another tool links to that vendor's own site or docs, checked November 13, 2026.

What are robotics data tools?

Robotics data tools, sometimes called robot data management software, work with the data robots and fleets record: camera video, LiDAR point clouds, joint states, actions, and other sensor streams on one clock. Teams use them across seven stages.
  1. Record and ingest. Log data on the robot and move it into storage.
  2. Replay and debug. Play a recording back, live or later, to see what the robot did.
  3. Search. Find moments across many recordings by metadata, signal values, or what appears on camera.
  4. Curate. Choose the episodes worth training on, such as teleoperation demonstrations, edge cases, and failures, and group them into datasets.
  5. Annotate. Add labels such as 3D cuboids, segmentation masks, or tagged time ranges.
  6. Evaluate models. Measure how a model or policy performs and find where it fails.
  7. Hand off to training. Export datasets that training code reads, for VLA and imitation learning policies too, in formats such as LeRobot.
The last stages loop. Failures found in evaluation decide what to curate and label next. A robotics data platform covers several of these stages in one product. For the wider category, see physical AI data platform.
The seven stages of robotics data work. Failures found in evaluation decide what to curate next.

At a glance: robotics data tools compared

Foxglove and Rerun are built for recording and replay, Roboto for analyzing fleet logs, and Encord, Lightly, and Voxel51 for curation, labeling, and model evaluation. The groups overlap, and the two tables show where.
StageFoxgloveRerunRoboto
Record and ingest● Foxglove SDK in C++, Python, and Rust logs to MCAP. Auto-upload of MCAP and ROS 1 bag files● Open source logging SDK in C++, Python, and Rust● Ingests ROS, MCAP, PX4, ArduPilot, Parquet, CSV, and video logs
Replay and debug● Live and recorded, remote access, teleoperation, comparison mode● Open source viewer, live and recorded◐ Time series and video viewed together
Search● Structured search across recordings. Semantic search (beta, by request)● SQL and dataframe queries, local or in Rerun Hub● Tags, metadata, topic statistics, signal similarity
Curate● Versioned datasets of episodes (beta), plus sessions, events, and an agent (beta)◐ Dataset mixes as queries in Rerun Hub● Versioned Collections. Agents curate moments into Collections
Annotate◐ Events on time ranges◐ Annotation in the Rerun Hub viewer, and annotations as columns in the SDK◐ Events. Labeling integration on Enterprise
Evaluate models○ No model metrics listed○ No model metrics listed◐ Actions test algorithms and generate reports
Hand off to training● Datasets (beta) stream to PyTorch and Ray Data or download as MCAP. Exports MCAP and ROS 1 bag● LeRobot export, PyTorch dataloader (experimental)● Exports event Collections to LeRobot v3.0 or v2.1
Tools built around the recording, by stage. ● core feature, ◐ covers part of the stage, ○ not listed on the vendor's site or in its docs. Checked November 13, 2026.
StageEncordLightlyVoxel51
Record and ingest◐ Managed data collection service◐ Data collection service. Indexes MCAP with its SDK◐ Imports MCAP and LeRobot. No on-robot logging
Replay and debug◐ LiDAR, camera, and radar on one timeline. MCAP in beta◐ MCAP point-cloud sequences◐ Camera, LiDAR, and sensor streams in sync on one timeline
Search● Natural-language and similarity search, plus 40+ data metrics● Text and similarity search, object-level embeddings● Natural-language, similarity, and event search. MCAP indexing and segment embeddings (beta)
Curate● Model-in-the-loop curation● Sampling for diversity, deduplication, and class balance● Deduplication, smart data selection, episode-level curation
Annotate● Images, video, and LiDAR with SAM 2, SAM 3, tracking, and consensus workflows. MCAP in beta. Managed labeling● 2D labels and 3D cuboids on MCAP point clouds, plus labeling services● 2D and 3D labels on all data types on every plan. Auto-labeling on Enterprise or as an add-on
Evaluate models● Rubric-based and pairwise evaluation, label error detection, model analytics● Confusion matrix and per-sample metrics● Model comparison and scenario evaluation, from metrics to samples
Hand off to training● Labels export as JSON, COCO, or YOLO● LightlyTrain. COCO, YOLO, Pascal VOC, YouTube-VIS, and MCAP sequence export● LeRobot v3 export
Tools built around the dataset and the model, by stage. ● core feature, ◐ covers part of the stage, ○ not listed on the vendor's site or in its docs. Checked November 13, 2026.
How to read these tables. ● the vendor offers this as a core feature. ◐ it covers part of the stage. For hand off to training, ● means the tool exports a format training code reads directly or streams data into training. ○ the vendor's site and docs don't list it. Facts come from each vendor's own site and docs, checked November 13, 2026, with sources in each tool's section below. We make Voxel51.
Best for, in one line each:
  • Foxglove: teams that need to get fleet data off robots, watch them live, and debug runs.
  • Rerun: engineers who log and visualize multimodal data from code with an open source SDK.
  • Roboto: teams analyzing signals and events across large volumes of robot logs.
  • Encord: teams that want labeling software plus managed labeling and data collection.
  • Lightly: teams pairing curation and labeling with model pretraining in LightlyTrain.
  • Voxel51: teams that curate, label, and evaluate MCAP and LeRobot recordings, with LeRobot v3 import and export.

Robotics data guides from Voxel51

How do you choose robotics data tools?

Start from the job you need done. Each answer below names the tools built for that job, in the order data moves through the stages.

Which tools record and ingest robot data?

Foxglove and Rerun record data on the robot, and Foxglove and Roboto move recorded logs into storage. Foxglove records MCAP with its SDK and auto-uploads MCAP and ROS 1 bag files. Rerun logs from C++, Python, and Rust into its own .rrd format. Roboto ingests ROS, MCAP, PX4, ArduPilot, and other logs. Encord and Lightly also sell data collection as a service. Voxel51 starts after recording, from MCAP or LeRobot files.

Which tools replay and debug a robot run?

Foxglove and Rerun are built for replaying and debugging a single robot run. Foxglove plays data live or recorded, with remote access, teleoperation, and side-by-side run comparison. Rerun's open source viewer shows live or recorded data, with views you configure in code or in the app. Roboto, Encord, Lightly, and Voxel51 also play back recordings, but debugging a running robot isn't their focus.

Which tools search ROS bag and MCAP recordings for failures and edge cases?

All six tools search robot data, in different ways. Foxglove, Rerun, Roboto, and Voxel51 search inside robot recordings for failure episodes and edge cases. Foxglove searches message content, events, and metadata, with semantic search over video in beta and by request. Rerun queries recordings with SQL. Roboto finds similar signals. Voxel51 searches inside episodes by metadata, events, labels, and natural language, with MCAP indexing and segment embeddings in beta for finding similar moments. Foxglove and Roboto search ROS 1 bags directly. Rerun and Voxel51 search MCAP, so ROS bags are converted first. Encord and Lightly search by text and similarity, with MCAP support in beta at Encord and MCAP point-cloud sequences at Lightly.

Which tools curate robot training data?

Foxglove, Roboto, Encord, Lightly, and Voxel51 curate robot training data, and Rerun Hub defines dataset mixes as queries. Foxglove builds versioned datasets of episodes, in beta. Roboto groups data into versioned Collections and exports event Collections to LeRobot. Encord curates with a model in the loop. Lightly samples for diversity, deduplication, and class balance. Voxel51 curates MCAP and LeRobot episodes and exports LeRobot v3.

Which tools label robot and sensor data?

Encord, Lightly, and Voxel51 label robot and sensor data. Encord labels images, video, and LiDAR with SAM 2, SAM 3, tracking, and consensus workflows, supports MCAP in beta, and sells managed labeling. Lightly labels 3D cuboids on MCAP point clouds and sells labeling services. Voxel51 labels 2D and 3D data, including 3D cuboids and polylines synced to camera views, on every plan, with auto-labeling on Enterprise or as an add-on. Foxglove and Roboto mark events on time ranges, and Rerun stores annotations alongside recordings. More in the annotation guide.

Which tools evaluate robot models and policies?

Encord, Lightly, and Voxel51 evaluate models. Encord runs rubric-based and pairwise evaluation and detects label errors. Lightly reports a confusion matrix and per-sample metrics for detection, classification, and segmentation. Voxel51 compares models across scenarios and data slices, then sends failure cases back to curation and labeling. Roboto actions can test algorithms and generate reports. More in the model evaluation guide.

Tools built around the recording

Foxglove, Rerun, and Roboto start where the data starts: on the robot. They record, store, replay, and search logs, and some now curate too.

Foxglove

Foxglove is a robotics data platform for collecting, visualizing, and debugging fleet data, with structured search across recordings and semantic search in beta, by request. It builds versioned datasets of episodes (beta), and its August 2026 launch added an agent (beta) that can add similar episodes to a dataset.
  • Best for: teams that need to get fleet data off robots, watch them live, and debug runs.
  • Robot formats: MCAP and ROS 1 bag in the data platform. ULog and ROS 2 bag files open locally.
  • Free option: free plan. The MCAP library is open source (MIT).
  • Runs in your environment: bring your own storage on AWS, Google Cloud, or Azure, an on-prem data plane in your own Kubernetes, or fully offline by contacting Foxglove.
  • Sources: foxglove.dev, pricing, August 18, 2026 launch post, data docs, Datasets, search docs, SDK.
  • Compare: Voxel51 vs Foxglove.

Rerun

Rerun is an open source SDK and viewer (Apache-2.0 and MIT) for logging, querying, and replaying multimodal data, with a commercial Rerun Hub for teams. The free SDK includes SQL and dataframe queries over a local catalog.
  • Best for: engineers who log and visualize multimodal data from code.
  • Robot formats: its own .rrd format. Opens MCAP and converts it to .rrd. ROS bags must be converted to MCAP first. Reads LeRobot v2.1 and v3.0 datasets.
  • Free option: the SDK and viewer are free and open source. Rerun Hub is priced through sales.
  • Runs in your environment: the SDK and local catalog run anywhere. Rerun Hub is a single-tenant deployment run for you, with data in your own S3-compatible buckets. For large teams, Rerun can deploy the data plane into your own account.
  • Sources: rerun.io, pricing, Rerun Hub, MCAP docs, reading LeRobot, LeRobot export docs, training docs.
  • Compare: Voxel51 vs Rerun.

Roboto

Roboto describes itself as an analytics engine for physical AI. It ingests robot logs in many formats, searches signals and events across them, and groups data into versioned Collections for training.
  • Best for: teams analyzing signals and events across large volumes of robot logs.
  • Robot formats: ROS, MCAP, PX4, ArduPilot, Parquet, CSV, and video, through ingestion actions.
  • Free option: free plan for up to 3 users and 100GB.
  • Runs in your environment: connect your own storage bucket on Premium. Self-managed deployment on Enterprise.
  • Sources: roboto.ai/platform, pricing, docs, Collections, formats, convert to LeRobot, agent curation.

Tools built around the dataset and the model

Encord, Lightly, and Voxel51 start from the training set. They curate, label, and evaluate, and two of them also sell data collection and labeling services.

Encord

Encord describes itself as the multimodal data layer for physical AI. It pairs annotation, curation, and model evaluation software with managed labeling and data collection.
  • Best for: teams that want labeling software plus managed labeling and data collection.
  • Robot formats: MCAP in beta, with 3D cuboids, polylines, keypoints, and point cloud segmentation, enabled by request. Encord's docs also list ROS bag (.bag) and ROS 2 (.db3) files as point cloud formats.
  • Free option: none listed. Starter, Team, and Enterprise plans, with prices not published.
  • Runs in your environment: API-first, with data kept in your cloud. VPC and on-prem deployment are Enterprise add-ons.
  • Sources: encord.com, MCAP in Encord, MCAP docs, label export docs, curation, pricing, supported data, October 2026 update, LiDAR.
  • Compare: Voxel51 vs Encord.

Lightly

Lightly describes itself as the data platform for physical AI. LightlyStudio curates and labels data, LightlyTrain pretrains and fine-tunes models, and Lightly sells data collection and labeling services.
  • Best for: teams pairing curation and labeling with model pretraining.
  • Robot formats: MCAP point-cloud sequences and 3D cuboid labels, added in LightlyStudio 1.1.3 and 1.2.0 (October 2026). Version 1.2.1 (November 2026) exports MCAP sequences and writes cuboid labels as MCAP.
  • Free option: open source LightlyStudio (Apache-2.0) for one user and one dataset, plus a 14-day trial of the hosted version.
  • Runs in your environment: the open source edition runs locally. Enterprise is hosted by Lightly or on-prem, including air-gapped.
  • Sources: lightly.ai, LightlyStudio, LightlyStudio changelog, Enterprise, on-premise, export docs.
  • Compare: Voxel51 vs Lightly.

Voxel51

Voxel51 is a multimodal data platform for visual and physical AI. It searches, curates, labels, and evaluates images, video, LiDAR, and robot recordings in one place, and reads MCAP and LeRobot natively on every plan. MCAP indexing and segment embeddings are in beta. We make Voxel51.
  • Best for: teams that curate, label, and evaluate MCAP and LeRobot recordings.
  • Robot formats: MCAP and LeRobot v3, with camera, LiDAR, and sensor streams in sync on one timeline. ROS bags must be converted to MCAP first.
  • Free option: free plan with 3 datasets of up to 10,000 samples each, 20GB of storage, and 30 hours of compute a month, with no time limit.
  • Runs in your environment: self-hosted on Team and Enterprise. On-prem and fully air-gapped on Enterprise.
  • Sources: pricing, robotics.

How do teams combine robotics data tools?

Most teams use more than one tool and pass files between them. Two open formats make that work: MCAP for recordings and LeRobot for robot learning datasets. None of the paths below is a built-in integration. They work because both tools read or write the same file. Encord and Lightly can also take MCAP from recording tools, Encord in beta and Lightly for point-cloud sequences.
  1. Foxglove, then Voxel51. Record and upload MCAP with Foxglove and debug runs there. Export MCAP for the episodes that matter, then search, curate, label, and evaluate them in Voxel51. Export LeRobot v3 for training.
  2. Rerun, then LeRobot. Log with the Rerun SDK, query recordings in a catalog, then export LeRobot datasets with the rerun-lerobot package.
  3. Roboto, then LeRobot. Ingest fleet logs into Roboto, group the moments you want into an event Collection, and export it as a LeRobot dataset. The export writes LeRobot v3.0 by default, with v2.1 as an option. Voxel51 reads LeRobot v3, so the exported episodes can be curated, labeled, and evaluated there.
ROS teams usually convert bag files to MCAP first. MCAP has been the default ROS 2 storage format since the Iron release, and most tools in this guide read it.

Which robot data formats does each tool support?

All six read MCAP in some form, Encord in beta and Lightly for point-cloud sequences. Foxglove, Roboto, and Encord also list ROS 1 bags, and Roboto and Encord list ROS 2 bags too. Rerun and Voxel51 take ROS bags after conversion to MCAP. LeRobot export is available from Rerun, Roboto, and Voxel51, and Rerun and Voxel51 also read LeRobot datasets. For PX4 ULog files, Roboto ingests them and Foxglove opens them locally.
ToolMCAPROS 1 bagROS 2 bag (.db3)LeRobotExports
FoxgloveReadsReadsLocal files onlyReplays through LeRobot's dataset viewerMCAP, ROS 1 bag, CSV of search results. Datasets (beta) stream to PyTorch and Ray Data
RerunReads, converts to .rrdConvert to MCAPConvert to MCAPReads v2.1 and v3.0. Exports with rerun-lerobot.rrd, LeRobot
RobotoReadsReadsReadsExports v3.0 or v2.1LeRobot
EncordReads (beta, by request)Reads point cloudsReads point cloudsNot listedLabels as JSON, COCO, or YOLO. MCAP labels through the SDK
LightlyReads point-cloud sequencesNot listedNot listedNot listedCOCO, YOLO, Pascal VOC, YouTube-VIS, MCAP sequences
Voxel51ReadsConvert to MCAPConvert to MCAPReads and exports v3LeRobot v3
Robot data formats each tool reads and exports, from each vendor's docs, checked November 13, 2026.

Which robotics data tools have a free plan or run in your own environment?

Foxglove, Roboto, and Voxel51 have free hosted plans. Rerun and Lightly offer free open source editions. Encord lists no free option. All six can keep data in your own infrastructure on some plan.
ToolFree optionRuns in your environment
FoxgloveFree planYour own storage on AWS, Google Cloud, or Azure. On-prem data plane in your Kubernetes. Fully offline by contacting Foxglove
RerunOpen source SDK and viewer. Rerun Hub priced through salesSDK and local catalog run anywhere. Rerun Hub is single-tenant, with data in your S3-compatible buckets
RobotoFree plan: up to 3 users, 100GBYour own bucket on Premium. Self-managed on Enterprise
EncordNone listed. Prices not publishedData kept in your cloud. VPC and on-prem as Enterprise add-ons
LightlyOpen source LightlyStudio for one user and one dataset. 14-day hosted trialOpen source runs locally. Enterprise on-prem, including air-gapped
Voxel51Free plan: 3 datasets, 10,000 samples each, 20GB, 30 compute hours a month, no time limitSelf-hosted on Team and Enterprise. On-prem and air-gapped on Enterprise
Free options and deployment for each tool, from each vendor's pricing and docs, checked November 13, 2026.

What about other robotics data tools?

This guide compares software teams use to turn recorded robot data into training data and model evaluations. These tools and categories come up in the same searches but do a different main job.
Fleet operations and telemetry
  • Formant describes itself as the operating layer for physical AI, with monitoring, intervention, and orchestration for robot fleets.
  • InOrbit is a robot orchestration platform for managing and monitoring robot fleets.
  • Sift is a hardware intelligence platform for telemetry, from development through live operations, in industries such as space, aviation, defense, and robotics.
  • Heex captures event-triggered data from robots and vehicles at the edge and routes it to the right teams.
Labeling-first platforms and services
  • Segments.ai is a multi-sensor labeling platform for robotics and autonomous vehicles, covering 2D images and 3D point clouds.
  • Avala captures and labels robot sensor recordings, with labeling software and managed teams. It ingests MCAP, imports ROS bag camera streams, and imports and exports LeRobot.
  • Scale collects and annotates physical AI data as a managed service, alongside its data platform.
Storage and research projects
  • ReductStore is a time-series database for storing and managing robot data.
  • ARES is an open source (Apache 2.0) robot data platform for researchers, published by a16z in April 2025.

How we compared these tools

  • Scope. Software that robotics and physical AI teams use on recorded robot data, with public product documentation. Data providers, fleet operations tools, storage, and research projects are listed separately above.
  • Sources. Each vendor's own website, docs, pricing page, and changelog, opened and checked November 13, 2026. Each tool's section links its sources.
  • "Not listed" means the vendor's site and docs don't state it. It doesn't mean the feature can't exist.
  • Beta marks features a vendor labels as beta, for every vendor, including us.
  • Disclosure. We make Voxel51. We don't rank tools, and we describe each one by the jobs its vendor says it does.

Try Voxel51 with your robot data

Upload MCAP or LeRobot episodes to a free Voxel51 workspace. Play camera, LiDAR, and sensor streams in sync, search inside episodes, label what matters, and see where your model fails. No credit card, no time limit. Get started.

Try Voxel51 for yourself

Frequently asked questions about robotics data tools

Have robot recordings to curate? Try Voxel51 for free · See Voxel51 for robotics · Book a demo