On October 19 Dream Labs came out of stealth with a $90M seed round and opened a public API for Dream-1. Dream-1 is a World Action Model (WAM) that learns to predict video and actions together. The company says it can zero-shot control bimanual parallel-gripper robots without any demonstrations from that robot, and it ran 27 unseen robots in simulation, including the 18-year-old PR2. Pricing is $3/hour for Dream-1 and $2/hour for Dream-1-flash. In an independent Robocurve evaluation, Dream-1 scored 59 on RoboDojo-RC Tier 1 versus 48 for the previous best. The company stresses it is a research beta not validated for safety-critical use.
Key Takeaways
- ✓Funding and team: $90M seed; the founders introduced World Action Models in the DreamZero paper in February 2026 while at NVIDIA.
- ✓Pricing: $3/hour for Dream-1, $2/hour for Dream-1-flash; each call returns 1.6 seconds of actions, about 2,250 calls per hour, roughly $0.0013 per call (company figures).
- ✓Generalization: zero-shot control of 27 unseen bimanual robots in simulation (including PR2); beta testers at 12 sites in 8 countries ran real-robot tasks.
- ✓Benchmark: Robocurve's independent eval gives 59 on RoboDojo-RC Tier 1 (previous best 48), 16x lower cost per attempt than Gemini 4 Argon and 1.7x faster than GPT-6 Astra; all 120 runs are public.
- ✓Limits: it predicts end-effector poses and relies on each robot's IK solver; robots far from its training can be adapted with the FaaS fine-tuning service using a little of their own data.
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the industry pain point
Robotics has long been stuck in a chicken-and-egg loop: a robot must be useful before people buy it, yet usefulness needs lots of data from that exact robot. Dream Labs argues two things are changing this: bimanual arm kits now cost around $2,000, and a model that generalizes zero-shot can make a new robot useful on day one. The founders introduced World Action Models (WAMs) in the DreamZero paper in February 2026 while at NVIDIA, and are now independent with a $90M seed.
Architecture and mechanism
Dream-1 watches the scene through three cameras (a head view plus one wrist view per arm) and learns to predict video and actions together. The company's reasoning is that, from the head camera, any robot is just shapes moving through a scene, so if the model can imagine the robot completing the task it can derive the actions. It can output joint angles, or end-effector poses and gripper state, an action space shared by every two-arm parallel-gripper robot, which each robot's IK solver converts into joint commands.
Benchmarks
Robocurve evaluated Dream-1 independently before launch: 59 on RoboDojo-RC Tier 1 versus 48 for the previous best, 16x lower cost per task attempt than Gemini 4 Argon and 1.7x faster than GPT-6 Astra, with all 120 runs and videos published. It ran 27 unseen robots zero-shot in simulation. These are company and evaluation-partner figures.
Getting started
Dream API is open to everyone, and the Python client takes a few lines: send camera images, robot state and a task, and get the next chunk of actions. Each call returns 1.6 seconds of actions, at $3/hour for Dream-1 and $2/hour for Dream-1-flash. Without a robot, use the Robot Playground to drive real robots in the SF office remotely, or pick a model in the Sim Playground. Robots far from its training can use the FaaS fine-tuning service. It is a research beta, so test in controlled settings and take extra care near people.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.