Case Study: Physical Intelligence runs real-time robot inference remotely on Modal
Key results
The challenge
Physical Intelligence builds general-purpose robotic systems that need continuous real-time inference across a fleet running 24/7. Local inference requires an expensive onboard GPU per robot, while conventional cloud inference adds latency and request-response bottlenecks incompatible with real-time control loops.
The solution
Physical Intelligence partnered with Modal to build a QUIC-based portal over UDP with automatic NAT traversal, replacing TCP request-response patterns. Persistent bidirectional channels between robot runtimes and Modal GPU containers enable continuous observation-to-action streaming without head-of-line blocking.
The results, in context
Physical Intelligence reported that the approach added only about 10-15ms of network latency for remote inference. Checkpoint loading completed in under 30 seconds and container startup reached service reachability in under a minute.