Evaluate & Improve Physical AI

Find your robot's failure modes in 30 minutes with GPU-accelerated simulation.

We’re getting AI into the physical world

01

Codifying the hardest tasks in simulation

Turn demanding physical tasks into high-fidelity, multi-modal simulation.

02

Evaluation Tasks defined by industry experts

To ensure reliability across complex industrial tasks, domain experts define scenarios and SOPs.

03

Making simulation accessible to all

Open tools and synthetic data so any team can train, test, and evaluate physical AI.

A living library of physical work, across every domain

• MARITIME
• SPACE
• AERIAL
• OFF-ROAD
• ROBOTICS
• OPERATIONS

The evaluation platform for robotics researchers

Run every major simulator benchmark in hours, not days, see exactly where your policy fails, and compare against verified baselines.

$ manifold run pi0.5 LIBERO-tiny  -> 1,000 episodes · 1h37m · score 0.72

Open source release coming soon

Evaluation today

• ✕A single eval rollout takes 24 hours or more
• ✕Every team reinvents the harness for every policy and every benchmark
• ✕No CI. Reproducibility is nearly impossible

1. Every benchmark LIBERO, RoboCasa, or your own scenarios
2. 8× faster sharded across GPUs by default
3. Track the SOTA compare against verified baselines
4. Find failures cluster episodes by failure mode

Synthetic data for perception and autonomy

Define a scenario in Python, or just describe it to your coding agent (Claude Code, Codex, any LLM). Stardust renders the rest: diverse, labeled, photorealistic data with rich time-series metadata, no 3D expertise required.

import bbi

world = bbi.World()
world.spawn("container_ship", quantity=12)
world.spawn("buoy", quantity=30, scatter=True)
world.ocean(sea_state=4)
world.weather(fog=0.3, time="dusk")
cam = world.camera(preset="maritime_eo")
imgs = world.render(frames=4000)
imgs.download(annotations=["bbox", "segmentation"])

Multi-domain, photorealistic data matching your operational requirements

One scene, rendered across modalities in perfect registration, with ground-truth labels generated automatically for every frame.

• RGB Photorealistic visible spectrum
• IR Infrared and thermal bands
• DEPTH Per-pixel depth and range
• SEGMENTATION + BBOXES Pixel-perfect masks and 2D/3D boxes

Realism and diversity with AI post-processing

A learned post-processing pass closes the domain gap and multiplies diversity. One scene, every condition: weather, lighting, and time of day, all photorealistic, so models transfer cleanly to the real world.

CLEAROVERCASTFOGDAWNDUSKRAIN

A 1,000+ object 3D asset library, and counting

Drag and drop from over a thousand production-ready, physically-accurate 3D assets, each tagged with real-world dimensions. Need something rare? Generate new assets with AI on demand, or request a custom asset from our 3D team.

Container ship 213 × 43 × 52 m

Container ship (Maersk) 400 × 60 × 73 m

Cruise ship 366 × 69 × 78 m

Response boat 14 × 4.7 × 8.1 m

Kayak 5.5 × 2.1 × 0.5 m

Monitoring buoy 1.0 × 1.0 × 1.1 m

Special mark buoy 2.8 × 2.6 × 6.1 m

Cardinal mark buoy 1.3 × 1.3 × 3.0 m

Cessna 208 Caravan 13 × 16 × 4.6 m

Pilatus PC-12 14 × 16 × 4.2 m

Harbor crane 12 × 28 × 29 m

Shipping port 912 × 1979 × 82 m

Build physical AI at the speed of compute

From synthetic data to policy evaluation, Bifrost is the infrastructure for AI in the physical world. Tell us what you’re building.