Physical AI developer platform

Build on the infrastructure your Physical AI models learn from.

Connect your agents to the hosted MCP server, build on the Python and TypeScript SDKs, read the metadata on every sequence, and import or export LeRobot datasets.

Start Building Book a technical walkthrough

Hosted MCP

Query the fleet learning loop from your AI tools.

Connect one hosted endpoint. Your agents can read fleet recordings and alerts, sequences, capture campaigns and their acceptance coverage, consensus, quality targets and export status.

Collection briefs, gap reports and evaluation reports come with the fleet learning program. Your Avala team delivers them.

Read the MCP guide

Hosted MCP endpoint

https://mcp.avala.ai/mcp

Tools by loop stage

Observe: fleet_list_recordingsfleet_list_eventsfleet_list_alerts
Diagnose: list_sequencespreview_curation_candidates
Acquire: list_capture_campaignsget_campaign_acceptance_coverage
Validate: get_dataset_healthget_consensus_summarylist_quality_targets
Release: list_sliceslist_exportsget_export_status

Access follows your Avala permissions. Reviewed staff actions appear only for eligible operators and require explicit confirmation.

Interfaces

Choose the interface that fits your stack.

Python, TypeScript, the public REST API, and hosted MCP share the same Avala data model.

Python SDK

Sync and async clients, cursor pagination, typed models, and explicit errors.

pip install avala

client.fleet.recordings.list()
client.datasets.get_health(owner, slug)
client.exports.create(dataset=dataset_uid)

TypeScript SDK

Typed resource clients with zero runtime dependencies and formal support for Node.js 18+.

npm install @avala-ai/sdk

avala.fleet.recordings.list()
avala.datasets.getHealth(owner, slug)
avala.exports.create({ dataset })

REST API

Use the public api.avala.ai endpoint from any language or authenticated HTTP client.

curl --fail-with-body -H "X-Avala-Api-Key: $AVALA_API_KEY" "https://api.avala.ai/api/v1/datasets/?limit=10"

One schema on every sequence you receive.

Sequences delivered through a fleet learning program carry the same fields. Your training code can filter, split and weight them without guesswork.

See how each sequence is graded

Outcome label on both grades
Subtask boundaries with a language description of each step
Quality and speed grades
Mistake and recovery segments
Rig and calibration IDs
Leakage groups by home, object and operator

Leakage groups let you split training and test sets so that the same home, object or operator never appears on both sides.

LeRobot

Move data in and out in LeRobot format.

Import a LeRobot dataset into Avala as MCAP sequences, or export an Avala sequence dataset as a LeRobot v3 dataset. Each LeRobot episode maps to one Avala sequence. Both commands ship with the Python SDK and need Python 3.12 or later.

Install the CLI with LeRobot support: Install the CLI with LeRobot support
pip install "avala[cli,lerobot]"
Import a dataset from the Hugging Face Hub: Import a dataset from the Hugging Face Hub
avala import lerobot --repo-id lerobot/svla_so101_pickplace --name "SO-101 pick and place" --slug so101-pick-and-place
Export a sequence dataset to LeRobot v3: Export a sequence dataset to LeRobot v3
avala lerobot export my-org/so101-pick-and-place --repo-id my-hf-user/so101-pick-and-place --output ./lerobot-out

First loop

Inspect your data with one authenticated request.

Enter your key through a no-echo prompt and query the public API while the credential stays isolated inside a temporary shell.

Bash or Zsh · curl · an Avala API key

Authenticated public API · Bash or Zsh

(
  printf "Avala API key: " &&
    IFS= read -rs AVALA_API_KEY &&
    printf "\n" &&
    export AVALA_API_KEY &&
    curl \
      --fail \
      --silent \
      --show-error \
      -H "X-Avala-Api-Key: $AVALA_API_KEY" \
      "https://api.avala.ai/api/v1/datasets/?limit=10"
)

Programmable workflow

Program the full Physical AI data loop.

Use stable interfaces for the operational steps between raw sensor data and the next model run.

Ingest sensor data

Bring MCAP, LiDAR, video, ROS bags, or LeRobot data into one dataset model.

CLIPython SDKREST API

Inspect dataset health

Check indexing, sequences, frame counts, and calibration before data enters production.

CLIPython SDKREST API

Run model-assisted labeling

Launch project-scoped auto-label jobs and keep human review inside the same governed workflow.

CLIPython SDKMission Control

Publish training-ready exports

Create a dataset or project export and track it through completion for downstream training.

CLIPython SDKREST API

Technical walkthrough

Bring us your hardest data loop.

We will map the interfaces, storage path, quality gates, and release contract around one real workflow.

Book a technical walkthrough Read the API reference