ARC-AI: Embodied Intelligence Platform
Physics-grounded autonomous robotic manipulation system. Foundation model-driven architecture with safety-first design.
Architecture
- Physics Encoder: Transformer (4 layers, 8 heads, 256 dim) pretrained on THE WELL
- Diffusion Policy: DDPM (1.5M params, 10-step DDIM inference)
- ACT Policy: Action Chunking Transformer (2.2M params, CVAE)
- Safety Layer: Rust, 500Hz real-time monitoring
- Control Loop: 1kHz position control (Franka 7-DOF)
Benchmark Results (A100-SXM4-80GB)
| Metric | Value |
|---|---|
| GPU Throughput | 8.97M samples/sec |
| Parallel Environments | 131,072 |
| Physics Step Rate | 32,667 Hz |
| Training Speed | 64 steps/sec |
| Adversarial Scenarios | 24 tested |
| Memory Efficiency | 1.21 GB peak |
Models
- Diffusion Policy — 500K steps, loss 0.046
- ACT Policy — 100K steps, loss 0.00011
Datasets
- Sim Demonstrations — 10K trajectories, 2.84M pairs
- SOTA Benchmarks — Full results + reports
Pretraining
Uses THE WELL (15TB physics simulations) for spatiotemporal pretraining before robot fine-tuning.
Tech Stack
Python 3.11+ | PyTorch | MuJoCo | Rust (safety) | Go (infrastructure) Zenoh (zero-copy IPC) | NATS JetStream | TensorRT | k3s
License
Apache 2.0