# Zero-Shot Robot Control: How Dream-1 Generalizes Across Unseen Hardware > Discover how Dream Labs uses video prediction and World Action Models to control unseen bimanual robots zero-shot over an API. **Author:** Drift Team **Published:** 2026-10-21 **Tag:** Research **Reading time:** 4 min **Canonical URL:** https://godrift.ai/blogs/dream-1 ---Training a robot policy traditionally requires collecting hours of teleoperation data on a specific physical arm before the machine can execute tasks reliably. On October 19, 2026, San Francisco-based startup Dream Labs, founded by former NVIDIA robotics researchers, emerged from stealth with a $90M seed round to address this bottleneck. The company introduced Dream-1, a World Action Model (WAM) accessed via public API that aims to control bimanual robots out-of-the-box without hardware-specific retrain cycles or fine-tuning.  
## **World Action Models and Video as a Unified Interface** Most Vision-Language-Action (VLA) models generate motor predictions directly from image inputs and language instructions. However, these representations often remain tightly coupled to the specific motor joint configurations of the training hardware. Dream-1 uses a World Action Model (WAM) architecture built on a pretrained video diffusion backbone. By treating video as a universal interface, the network predicts future visual frames alongside the physical actions required to realize those frames. Instead of outputting joint torques or motor voltage steps, Dream-1 predicts 6-DOF end-effector pose trajectories (where each gripper should navigate in 3D space) and gripper states (open or closed).   Because every two-armed robot equipped with parallel grippers shares this spatial end-effector action space, a single model can issue spatial target poses. The robot’s native Inverse Kinematics (IK) solver then translates those spatial targets into machine-specific joint commands. ## **Zero-Shot Transfer Across Bimanual Form Factors** To test cross-embodiment generalization, Dream Labs evaluated Dream-1 across 27 simulated robot models and over 20 real-world bimanual configurations—ranging from modern dual-arm setups to legacy platforms like Willow Garage's 18-year-old PR2.   In tabletop tests covering picking, placing, and folding tasks, Dream-1 demonstrated zero-shot task execution on unseen bimanual architectures without requiring prior demonstration data for those specific frames. The system accepts camera images, current robot state vectors, and plain-English prompts, streaming back action chunks every 1.6 seconds to maintain a closed-loop control cycle. ## **Operational Trade-Offs and Technical Limits** While zero-shot transfer reduces setup time for new hardware, the architecture introduces specific engineering considerations:   **IK Solver Dependence:** Because Dream-1 predicts target end-effector poses rather than joint torques, final motion execution depends on the precision and stability of the target robot’s local IK solver.   **Trajectory vs. End-State Control:** The model predicts where the end-effector should be positioned, but not always the exact sweeping arm motion between states. Path planning near strict physical obstacles or joint limits still requires local motion constraints.   **Kinematic Scope:** Dream-1 currently targets dual-arm setups with parallel grippers doing tabletop manipulation. Highly compliant systems, dexterous multi-fingered hands, or non-standard joint topologies require dedicated fine-tuning or specialized model adaptations. ## **Reducing Barriers to Embodied AI** Delivering robot control policies via a hosted API at $3 per hour shifts the primary bottleneck in robotics from model architecture design to hardware accessibility. By combining video prediction with end-effector pose outputs, World Action Models offer a viable path toward general-purpose robot brains that can control diverse physical systems without retraining from scratch. ## **FAQ** ### **What is a World Action Model (WAM)?** A World Action Model predicts visual video frames and physical action trajectories together. By predicting how a physical scene evolves frame-by-frame, the model imagines future states and calculates the spatial movements required to achieve those states. ### **How does Dream-1 achieve zero-shot transfer across different robots?** Dream-1 predicts 6-DOF end-effector (gripper) poses rather than raw joint commands. Because bimanual robots with parallel grippers share an end-effector pose space, the target robot's internal Inverse Kinematics (IK) solver handles the translation from spatial poses to motor movements. ### **What hardware configurations does Dream-1 support?** Dream-1 currently supports tabletop bimanual (two-armed) robots equipped with parallel grippers. It is designed for manipulation tasks like picking, placing, sorting, and folding. ### **Is zero-shot control as reliable as fine-tuned control?** Zero-shot control allows a robot to begin executing tasks immediately on unseen hardware. However, for specialized, highly precision-critical, or contact-rich industrial tasks, fine-tuning the model on robot-specific demonstrations yields higher task success rates. ## **Related Reading** - [Why Robots Need to Understand Space, Not Just Objects](https://www.godrift.ai/blogs/why-robots-need-to-understand-space) - [5 Humanoid Robots Pushing Robotics in New Directions](https://www.godrift.ai/blogs/5-humanoid-robots-pushing-robotics) - [PID vs. MPC vs. Reinforcement Learning: How Robots Control Movement](https://www.google.com/search?q=https://www.godrift.ai/blogs/pid-vs-mpc-vs-reinforcement-learning-robot-control&authuser=3) - [AMD Just Bet $8.2 Billion on Physical AI](https://www.godrift.ai/blogs/amd-just-bet-8-billion-on-physical-ai)