EulerAtlas

Plugin-based robot behavioral-learning framework — first public: Car · Humanoid

A plugin-based CLI robot behavioral learning framework running on an RL→IL pipeline that combines imitation learning (IL) with FastTD3 reinforcement learning (RL). The first two flagship domains we are releasing are car (EulerDrive) and humanoid (EulerWalk) — EulerDrive delivers multi-camera BEV autonomous driving (closed-loop verified on CARLA Town01), and EulerWalk delivers a humanoid RL→IL loco-manipulation policy, with both bundling pretrained models in the package so you can load and run them with no training or simulator at all. The unified obs/action schema is designed to cover 8 domains — car, drone, humanoid, robot dog, mobile manipulation, logistics AGV, agri-robot, shipyard crane — and each domain ships and keeps expanding independently as a Domain Plugin.

Open Source · v0.5.0

Core Features

First-public 2 domains (car · humanoid) + 8-domain schema expanding via plugins · IL + FastTD3 RL · SOTA policy models

Classic Domains (01-04)

Internal Schema Boundary — all modules communicate consistently through domain-specific fixed-dimension schemas.

Car (Road) 4D state + 2D action — CARLA integration, nuScenes adapter
Drone (Aerial) 6D state + 3D action — AirSim skeleton
Humanoid 33D state + 14D action (loco-manip) — MuJoCo Humanoid-v5 / HumanoidBench integration
Robot Dog (Quadruped) 12D state + 8D action — MuJoCo Ant-v5 integration

Industrial Domains (05-08)

Factory, logistics, agriculture, shipyard — each domain includes a dedicated mock simulator and data adapter.

Mobile Manipulation 18D state + 8D action — Recommended model: ACT
Warehouse AGV 14D state + 4D action — Recommended model: BC-RNN
Smart Farm Agri-Robot 16D state + 6D action — Recommended model: Diffusion
Shipyard Crane 20D state + 9D action — Recommended model: ACT

Policy Models (core 5 + SOTA + RL)

BC-MLP / RNN / CNNBasic / time-series / vision-based behavioral cloning
ACTAction Chunking Transformer + CVAE (manipulation SOTA)
Diffusion PolicyDiffusion model-based policy (DDPM)
SOTA (plugin)BEVFuser (multi-camera BEV), TemporalTransformer (Tesla-FSD style), WholeBodyACT (separate loco/manip heads)
RL Actor NEWFastTD3 off-policy RL (twin critic, n-step) — RL→IL workflow

Complexity Levels (L0–L2)

L0 (Toy)Mock simulator + BC-MLP — for tutorials/CI
L1 (Intermediate)Domain-recommended model + Mock — algorithm research
L2 (Advanced)Real backends (CARLA / MuJoCo / HumanoidBench) + SOTA models

First-public flagships — EulerDrive (car) · EulerWalk (humanoid)

The first two domains we are releasing. Pretrained models ship in the package, so you can load and run policies with no training or simulator. The remaining domains roll out incrementally as plugins.

EulerDrive (car)

Multi-camera BEV-fusion autonomous driving. Built from BEVFuserPolicy + IPM BEV projection + a PID waypoint controller, verified closed-loop on CARLA Town01 3-camera data.

  • Closed-loop metrics: RouteCompletion / InfractionScore / DrivingScore.
  • Verified result: CARLA Town01 waypoint loss 14.71 → 4.56 (69% reduction).
  • Bundled model: bevfuser_lite_carla_town01.pt.

EulerWalk (humanoid)

A complete RL→IL pipeline. FastTD3 reinforcement learning (RLTrainer) builds an expert, generates demonstrations, then distills them into a WholeBodyACT imitation policy.

  • Verified result: RL return 406 → 418 → IL loss 1.98e-4, with stable rollouts.
  • Bundled models: wholebodyact_locomanip.pt, rl_actor_walk.pt.
  • Schema: humanoid 33D state / 14D action (loco-manipulation).

※ The 2 first-public domains (car/humanoid) are research-grade and built on real backends and datasets. The other 6 domains roll out as plugins on the same unified schema.

Data Pipeline & Simulation

Complete pipeline from synthetic data generation to simulation rollout

Data Pipeline

Perform the complete data workflow via CLI — from synthetic data generation and external data collection to augmentation and editing.

  • Pull: Generate/collect synthetic or real expert demonstration data.
  • Ingest: Convert JSONL/CSV to internal format.
  • Augment: Data augmentation with noise, mirroring, time jitter, etc.
  • Edit: Statistics review, filtering, merging, splitting.

Simulation & Evaluation

Run trained policies in simulation and automatically compute evaluation metrics.

  • Mock Simulator: 8 domain-dedicated built-in simulators — run immediately (no dependencies).
  • Real Backends: CARLA (cars), MuJoCo (robots), AirSim/IsaacLab (skeleton).
  • Parallel Simulation: Run N instances concurrently with ParallelSimManager.
  • Scenario Suite: Build corner-case scenarios and generate pass/fail reports.

CLI Reference

Command groups cover the entire workflow (--lang supports 5 languages)

init

Auto-generate YAML configuration from role and level.

train

Train BC/IL policies. Supports resume, GPU, checkpoints, and project tracking.

train-rl NEW

Train an expert policy with FastTD3 reinforcement learning (RL→IL workflow).

sim

rollout/replay — Run and evaluate policies. Supports --render / --save-video (MP4).

data

pull, ingest, augment, edit, collect, generate-demos — the data pipeline.

model

export/load — Manage model cards (SafeTensors + config.json).

skill

Query built-in skills, initialize, and perform cross-domain transfer.

deploy

Deploy policies via MessageBridge (mock/ROS2/ZeroMQ).

scenario

Build corner-case scenarios and run scenario suites.

validate

Validate YAML configuration files (50+ rules).

edge

Manage the edge training lifecycle — on-device training, deployment, monitoring.

plugin

Manage Domain Plugins — register, query, and initialize custom domains.

Architecture

A 4-layer abstraction stack separates roles, configuration, training, and evaluation

4-Layer Stack + Domain Plugin

Layer 4 Skill/Meta-Agent — Role/mission-based recommendations
Layer 3 Config/Validator — YAML manifest + 50+ validation rules
Layer 2 Data/Model/Training — Pipeline, policy networks, BCTrainer
Layer 1 Sim/Eval/Backend — Simulator, evaluation, trajectories
Domain Plugin Register custom domains (obs/action schema, simulator, adapters) as plugins

Technical Specifications

Language Python 3.11+
Framework PyTorch, Gymnasium
Simulators Mock (built-in), CARLA, MuJoCo
Logging W&B, MLflow (optional)
Model Format SafeTensors + config.json
Error Format 3-line format (Category / Fix / See)

Tutorials

Step-by-step guides to get started with EulerAtlas quickly

Tutorials coming soon.

Installation & Getting Started

Install EulerAtlas and train your first policy

Installation

pip install -e ".[dev]"

# Get started in 3 lines
euleratlas init --role road --level 0
euleratlas train -c config.yml
euleratlas sim rollout -c config.yml

Requirements

Python 3.11+, PyTorch

MuJoCo, W&B (optional)

Train Robot Policies Across 8 Domains with EulerAtlas

From cars to shipyard cranes, from expert demonstrations to policies in 3 CLI commands.

Get Started on GitHub Contact Us