Dream Labs Dream-1 release: zero‑shot robot control via public API

Dream‑1 delivers zero‑shot manipulation for unseen robots

Dream‑1, released by Dream Labs on 24 Oct 2026, is the first publicly available World Action Model (WAM) that can control a brand‑new bimanual robot with parallel grippers without any fine‑tuning or robot‑specific data. Users simply connect their robot to the Dream‑1 API, send an observation, and receive actions in real time. The model runs at $3 / hour (or $2 / hour for the lighter Dream‑1‑flash) and returns 1.6 seconds of motion per call, making robot control affordable and plug‑and‑play.

Why zero‑shot matters: breaking the robotics chicken‑and‑egg loop

Robotics has long suffered from a data‑cold‑start: a robot must be useful to generate data, but useful behavior requires data from that exact robot. Dream‑1 solves this by (1) leveraging cheap, mass‑market hardware – a capable two‑arm setup now costs ≈ $2 k – and (2) using a model that generalizes across embodiments, tasks, and environments without any robot‑specific training. This creates a virtuous flywheel: cheaper robots increase deployment, the model makes them useful immediately, and the resulting data further improves the model.

The hardware explosion that enables the flywheel

  • $2 k bimanual arm setups are now commercially available (e.g., MakerMods’ Maker Arm).
  • 7 semi‑humanoid robots under $15 k demonstrate that multi‑arm platforms are no longer enterprise‑only.
  • Hugging Face’s $399 Microduck sold >10 000 units in four days, showing strong consumer demand for affordable, cute robots.

These price points expand the potential robot user base from labs to hobbyists and small teams, providing the data volume needed for large‑scale model training.

Zero‑shot performance across 27 unseen robots

Dream‑1 was evaluated on 27 unseen robot embodiments in simulation and on 4 real robots (including the 18‑year‑old PR2). Relative to the YAM robot it was trained on, Dream‑1 achieved:

  • 4× faster execution on real robots and 3× faster in simulation.
  • 8× speedup for beta‑tester sites and 4× speedup in‑house compared to prior baselines.

These results constitute one of the first reported zero‑shot transfers to completely new robot bodies, confirming the claim that a WAM can predict both video and actions from pixel observations alone.

How Dream‑1 achieves embodiment‑agnostic control

  1. Shared end‑effector (EEF) action space – Dream‑1 predicts the 3‑D pose of each gripper and its open/close state. An off‑the‑shelf inverse‑kinematics (IK) solver translates these poses into joint commands for any bimanual parallel‑gripper robot.
  2. Pixel‑based world modeling – By learning to predict future video frames, the model treats every robot as a collection of moving pixels, regardless of its geometry. This removes the need for identical grippers or wrist‑camera‑only views, which limited earlier approaches such as the Universal Manipulation Interface (UMI).
  3. World Action Model training – Dream‑1 jointly learns video prediction and action generation, aligning perception and control so that imagined futures directly inform motor commands.

Limitations and the role of one‑shot fine‑tuning (FaaS)

Zero‑shot performance varies with how closely a new robot matches the training distribution. For robots with very different kinematics or gripper designs (e.g., Dexmate’s dexterous hands), Dream‑1 benefits from Fine‑tuning‑as‑a‑Service (FaaS):

  • 5 minutes of robot‑specific data can boost speed by ~3× on bimanual arms.
  • 77 minutes were needed for Dexmate’s hands, and 98 minutes for a compliant gripper on YAM, demonstrating that more divergent embodiments require more data.
  • Across twelve FaaS beta partners, language‑following scores improved on ten robots, confirming that one‑shot adaptation reliably narrows the zero‑shot gap.

Independent benchmark: Robocurve

Robocurve’s public evaluation placed Dream‑1 on a new Pareto frontier for cost‑per‑task and speed:

  • Mean score 59 on RoboDojo‑RC Tier 1, outperforming Opus 5.5 and Gemini 4 Argon.
  • 16× lower cost per task than Gemini 4 Argon and 1.7× faster than the fastest large language model (GPT‑6 Astra).

Community reaction on Hacker News

  • Positive curiosity – Users reported successful prompts such as “move the elephant next to the ball” and “put the crown on the cup,” confirming the API works for simple commands.
  • Skepticism about real‑world applicability – Several commenters (e.g., Aurornis, gradus_ad) argued that the demos feel toy‑like and doubt the platform’s relevance for industrial automation or manufacturing.
  • Concerns over control and safety – A user warned about third‑party control of home robots, while others questioned the broader economic impact of robots replacing manual labor.
  • Mixed impressions of the narrative – Some noted the announcement reads like AI‑generated marketing copy, making it harder to assess technical depth.

These viewpoints highlight both excitement about the technology and the need for clearer use‑case demonstrations beyond novelty tasks.

Getting started with Dream‑1

import dream
client = dream.DreamClient(api_key='<API_KEY>')
client.policy({"checkpoint": "dreamlabs/dream-1"})
client.set_observation(observation)
action = client.get_action()
client.close()

The above snippet shows the minimal code required to issue a policy request. Detailed setup instructions, including IK integration guides, are provided on the Dream Labs platform.

Outlook: from bimanual arms to general robot intelligence

Dream‑1 is positioned as a foundation model for manipulation. Dream Labs envisions extending the approach to mobile manipulators, dexterous hands, and entirely new morphologies. By exposing a low‑latency, cost‑effective API, they aim to democratize robot intelligence, allowing anyone with a compatible robot to experiment, collect data, and contribute to the next generation of general‑purpose robotics.

This article synthesizes the Dream‑1 release page and the top‑voted Hacker News comments, presenting a self‑contained technical overview suitable for citation.

Sources