Deploying a realtime robotics API: how Dream Labs got up to 35% faster round trips with fal’s World Model Accelerator

Industry:
Robotics and physical AI
Use-case:
Realtime cloud inference for a robot foundation model: robots stream what they see and get actions back
Results:
Dream API went live on fal Serverless 14 days after first contact, on reserved B200 GPUs across three US regions. In testing, fal's World Model Accelerator (WMA) cut robot round trip time with no change to the model or the GPU.

Dream Labs is a robotics foundation model lab in San Francisco, founded by researchers from NVIDIA's GEAR lab, the team behind GR00T and DreamZero. Its first model, Dream-1, controls robots it has never seen, on tasks it was never shown: it was tested on 27 robots it was never trained on, with beta testers in 9 countries. Dream Labs came out of stealth on October 19, 2026 with a $90M seed round led by B Capital, and launched Dream API the same day. Dream API lets anyone with a robot prompt Dream-1 remotely: stream what the robot sees, type a task, and get robot actions back, with no high-end GPU on the robot.

Customer Highlights
2 days
from first contact to Dream API running on fal
14 days
from first call to public launch, on guaranteed capacity
Up to 35%
faster robot round trip with fal WMA in testing, same model and same GPU
99.999%
of nearly 2 million inference calls served successfully, across three US regions

The Challenge: Serve a large-scale robot model in realtime, to robots anywhere, on a fixed launch date

When a robot runs on Dream API, it sends what its cameras see to Dream-1 in the cloud and waits for its next chunk of actions. Every millisecond in that loop is a slower robot.Dream-1 is built on a large-scale video generation model, and each copy runs on multiple GPUs with a fast interconnect. Dream API's users can be anywhere with an internet connection, starting with the US West, Central and East, so GPUs had to sit close to the robots calling them.Capacity was the other risk. Dream Labs' previous provider ran on spot capacity: a high GPU limit on paper, but no guarantee, and on some days fewer than 10 GPUs were available. With a public launch date already set for October 19 and demand unknown, an outage on day one was the outcome the team most wanted to avoid.

The Dream Labs team looked for an infrastructure partner that could:

  • Guarantee capacity on launch day: dedicated GPUs that are always on, with room to burst above them when demand arrives.
  • Keep round trip time low everywhere: GPUs in multiple regions and the shortest possible path from robot to model.
  • Migrate without a rebuild: run Dream Labs' existing container, replacing only the inference layer.
“For a robot foundation model, round trip time is the product. Before launching Dream API, we needed infrastructure that could guarantee capacity in every region from day one, not a GPU limit that might not hold when demand arrived.”
Scott Reed, Co-founder, Dream Labs

How Dream Labs and fal work together

The first test was the migration. Dream Labs' serving stack was already packaged as a container image, so the team pointed fal's Python SDK at that image and deployed it as a fal Serverless app. Dream API was running on fal two days after first contact, with no changes to the model or the serving code.

With fal, Dream API gets:

  • Guaranteed capacity across US regions. Dream Labs reserved B200 GPUs that are dedicated to Dream API and always on, split across US West, Central and East. Each region runs its own deployment, so every robot connects to a nearby GPU. When traffic exceeds the reservation, the same apps burst onto fal's shared fleet.
  • A direct line from robot to GPU with WMA. fal's World Model Accelerator opens a live WebRTC session between the client and the GPU running Dream-1. Camera frames stream in and actions stream back over that one connection, with no request gateway in the path. fal handles signaling, TURN servers and session lifecycle, so Dream Labs did not need to build any of it.
  • Same-day engineering support. A shared Slack channel opened on day one. fal's engineers and the WMA lead joined the channel, raised GPU limits for testing the same day, and answered every question within hours during the trial and the launch.

Outcome: Launched on time, on guaranteed capacity, with faster round trips on WMA

  • Live in two days, launched in fourteen. First contact on October 5, Dream API running on fal October 7, and public launch on October 19, with no launch-day capacity scramble.
  • Faster round trips with WMA. Measured per action from a robot on a home internet connection, against the standard HTTP API on the same model and GPU:
    Frame sizeStandard APIfal WMAChange
    320 × 1761.02 s0.73 s28% faster
    640 × 3601.55 s1.00 s35% faster
    The whole gain came from the network path. Inference held at 0.5 to 0.6 seconds on a B200, while transport dropped to about 170 ms at both resolutions.
  • Reliable from day one. 99.999% of nearly 2 million inference calls served successfully across three US regions.
“With fal's World Model Accelerator, we measured up to 35% faster round trips on the same model and hardware. This shows up in every robot action.”
Scott Reed, Co-founder, Dream Labs
Customer Highlights
Dream Labs
went from first contact to a public robotics API on fal Serverless in 14 days
Dream Labs runs its existing container on reserved GPUs across the US, so every robot connects to a nearby GPU. fal’s World Model Accelerator cut round trip time by up to 35% on the same model and GPU.
Bria needed to bring new models to market faster while expanding developer adoption
Bria partnered with fal in 2024 to make its models accessible to developers through fal's inference infrastructure and model marketplace. Today, Bria's engineers can take a model from research-ready to publicly available in days rather than weeks, without managing the underlying GPU infrastructure.
How VEED Brings Its Own Models to Developers with fal
VEED deploys its talking-video models as apps on fal Serverless, which runs them for the VEED editor and lists them in fal’s model gallery the day VEED announces them. Fourteen VEED endpoints are live on fal today.
AutoHDR
edits 25% of US real estate listings on fal Serverless.
AutoHDR’s editing models run on fal Serverless, where capacity follows the morning rush of shoots. Photographers get edited listing photos back in about 15 minutes, and a small team ships new versions as deployments, not infrastructure projects.

Ready to build with generative media?

Select option
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.