Build
Physical AI.

The development platform for Physical AI.From first run to production. Full lineage. No gaps.

The Problem

Physical AI Tooling Is Broken.

No lineage between your data, models, failures, and rollouts.

Evaluation is an afterthought bolted onto training pipelines.

Local dev and remote training are completely different worlds.

Debugging means stitching together disconnected tools.

Anveld Closes the Loop.

Local Development

RunSpec Parity

One frozen spec runs locally or on Kubernetes. No field changes, no drift.

Simulator-First

Bootstrap to your first replayable run in under 30 minutes on supported Linux.

Hot Reload

Iterate on models and policies without restarting your environment.

Offline-Capable

Work without connectivity. Artifacts spool locally and sync when reconnected.

Data & Lineage

Artifact Graph

Every model, config, dataset, and recording is versioned with full provenance.

Recordings

Multimodal capture with replay-oriented indexing. Seek within 2 seconds across 50 GB runs.

Dataset Slicing

Extract reusable data subsets from any recording with lineage preserved end to end.

Retention Policies

Ephemeral, standard, or durable retention classes for every artifact you produce.

Training & Evaluation

DAG Pipelines

Define multi-stage training pipelines in Python. Run locally or on Kubernetes with the same spec.

Scenario Engine

Versioned test cases with inputs, seed policies, and validators. Deterministic and reproducible.

EvalPacks

Bundle scenarios, metrics, thresholds, and baselines into repeatable evaluation suites.

Promotion Gates

Auditable promotion decisions with exact evidence linking model, config, eval, and actor.

Debugging & Observability

Unified Timeline

Align replay, logs, metrics, traces, and events on a single scrubable timeline.

Run Diffing

Compare any two runs or a candidate against a baseline side by side.

Failure Clustering

Group failures by tags and export them directly as training data or new scenarios.

Annotations

Comment on runs, tag artifacts, and share investigation context with your team.

Engineers First

Built for Engineers.

train.py
Python 3.13
01from anveld import train
02from anveld.config import load_config
03
04config = load_config("anveld.yaml")
05
06result = train(
07 config=config,
08 env="mujoco://franka-pick",
09 trainer="ppo",
10)
11
12result.evaluate(pack="pick-place-v2")
13result.promote(baseline="production")
Python SDK
CLI Toolchain
Local-First
Deterministic Replay
Plugin System
The Loop

One Continuous Loop.

RUN & CAPTURE
01
REPLAY & DEBUG
02
EXPORT & SLICE
03
TRAIN & EVALUATE
04
PROMOTE
05
Applications
01

Robotic Manipulation

02

Autonomous Vehicles

03

Aerial Systems

04

Industrial Automation

Your System Next