Dream Labs Puts a Zero-Shot Robot Policy Behind a $3-an-Hour API
Dream Labs claims zero-shot control of 27 unseen robots in simulation, but real-robot tests used earlier checkpoints and one-shot adaptation still beats zero-shot.
Dream Labs emerged from stealth with $90M in seed funding and opened a public API for Dream-1, a World Action Model it says can control unseen bimanual, parallel-gripper robots zero-shot. The model can output joint angles for embodiments it supports natively and also predicts end-effector poses, which each robot's inverse-kinematics solver turns into joint commands.
Dream-1 was trained on the YAM robot and reads a head camera plus a wrist camera on each arm. Because a World Action Model predicts video and actions together, any gripper appears as pixels, which the company says enables transfer to unseen bodies.
Dream Labs reports tests on 27 unseen robots in simulation and 4 real robots.
Key takeaways
- Simulated evaluations cover 27 unseen robots at 300 rollouts per embodiment: 150 rollouts across 50 tasks for language following and 150 rollouts across 150 tasks for unseen-task generalization. Real-robot tests used earlier checkpoints, and OpenYAM was tested on unseen tasks only.
- Robocurve, a public benefit corporation, evaluated Dream-1 on RoboDojo-RC Tier 1 and reported it above Opus 5.5 and Gemini 4 Argon, at 16x lower cost per task attempt than Gemini 4 Argon and 1.7x faster than GPT-6 Astra.
- API pricing is $3 per hour for Dream-1 and $2 per hour for Dream-1-flash. Each call returns 1.6 seconds of actions, so an hour is about 2,250 calls, or roughly $0.0013 per action.
- Fine-tuning-as-a-Service adapts the model with a partner's own data: five minutes on two partner robots, 77 minutes for Dexmate's dexterous hands and 98 minutes for YAM with swapped end effectors. Language-following improved on 10 of 12 beta robots.
What people think
Analysis details
7 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 7 of 8 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
Front-page placement invites astroturfing suspicion
Dream Labs had almost no public history when the submission reached the front page before any comment, prompting the claim that PR agencies use fake accounts to drive trending. The model was found through the Robocurve benchmark after a social-media post flagged Kimi K3.1 Pro's poor score, and HN strips Introducing from titles. The suspicion remains unresolved.
Third-party control of home robots draws objection
A public API that lets a third party drive a bimanual robot inside a home is treated as a reason to slow adoption, not accelerate it, with the argument that economies of scale carry negative externalities. An open-source release from a Chinese lab is offered as the preferred route, which would move control in-house without removing the capability.
Cheap robot labor reframes household economics
Robot labor priced at $3 an hour undercuts human wages, and one argument asks how a worker competes for housing against that rate. The counter is that automating chores could cut drudgery and reduce the need to keep part of the workforce in low-status jobs. Existing free house-cleaning services already place cameras in customers' homes, and customers tolerate the surveillance.
Public evidence rests on one playground prompt
The only hands-on check reported is a single playground prompt, Move the elephant next to the ball, with a link to the generated recording. A single self-reported run cannot establish general manipulation ability, and the benchmark that surfaced the model was itself found through a social-media post about a rival's poor score.