Fast robot control requires understanding how the scene is changing. A single image can reveal an object’s position but leave its motion and interaction progress ambiguous. World-action models (WAMs) bring video prediction to closed-loop control, making visual history a natural resource for action. Yet processing more history can delay the response it is meant to improve. This work asks how WAM control scales with visual context, and how those gains can be retained under real-time constraints.
Scaling WAM Context from Autoregressive Video Generation
Long-WAM connects context scaling to the ability to predict physical evolution from history. It first learns robot motion and interaction dynamics through long-sequence autoregressive video pretraining, then transfers this predictive foundation to action generation while preserving its causal temporal structure. The resulting policy conditions actions on both observed history and anticipated futures, making prediction an intermediate representation for control. The study then examines how the benefit of additional history depends on the video initialization, distinguishing access to a longer context from the learned ability to exploit it.
Context Scaling of World-Action Models
The duration of real observations in the causal prefix is scaled while the visual forecast and action horizon remain fixed within each benchmark. This tests whether additional past evidence improves the same near-term control decision, rather than changing how far the model predicts or acts.
Each window uses a separately trained model evaluated at its training context length; zero history retains the current observation. When early-episode history is missing, the initial frame is repeated.
Efficient Edge Deployment
Local inference avoids network delays, but video-first prediction is costly on limited onboard compute. Long-WAM is accelerated on NVIDIA GeForce RTX 5090, DGX Spark, and Jetson AGX Thor to support short-overlap execution while preserving visual imagination. Shared optimizations reduce common computation and data-movement costs; device-specific tuning addresses platform constraints, balancing portability and hardware efficiency.
The implementation preserves joint attention without a concatenated KV buffer. It also coalesces memory accesses, fuses scale/bias and cast operations, and simplifies VAE layouts and padding, subject to shape and backend constraints.
Implementation Details
Pretraining LongLive2.0-Robot uses approximately 30,720 GPU-hours on 64 NVIDIA H100 GPUs; downstream WAMs train on 16 GB200 GPUs. WAM training uses AdamW with a peak learning rate of an unspecified value, cosine decay, 5% warmup, BF16 precision, and gradient clipping at 1.0.
Robot experiments use a Unitree G1 humanoid and a YAM bimanual manipulator. The paper’s Appendix D lists baseline references.
Atomic Execution and Compositional Planning
Long-WAM alone achieves 67.9% Atomic-Seen success, compared with 39.6% for an unspecified method and 31.5% for GPT-6 Astra alone, providing a strong physical skill foundation. Querying the same policy every 15 steps yields 84.4% without a planner, close to the hierarchical system’s 85.6%.
Real-World Deployment
Dynamic manipulation on Unitree G1 and long-horizon tasks on YAM are evaluated with 20 trials per policy and condition.
Pretraining Data
The pretraining corpus contains 2,294,889 model-ready text-and-image-to-video samples from five dataset families. The paper reports six source subsets, including two AgiBot World subsets. The training, validation, and test splits contain 2,251,021, 21,931, and 21,937 samples, respectively.
The approximately 10,000-hour training scale is a window-equivalent estimate: assigning 16 seconds to each training sample yields 10,004.5 aggregate hours. It is not a measurement of unique raw footage; overlapping windows, source clip lengths, and padding affect that distinction.
Pretraining uses video prediction without action supervision, so sources need not share action coordinates or actuator dimensions. Action-space-independent supervision is a property of this stage; improved transfer to an unseen embodiment would require a separate evaluation.
Dynamic Composite Manipulation
The paper describes dynamic cup stacking on Unitree G1. Long-WAM coordinates grasping the blue cup, intercepting the moving green cup, and nesting it inside the blue cup. The baseline rollouts instead miss the green cup during interception.
Long-Horizon Manipulation
Long-WAM executes three YAM tasks that require successive object interactions. The sequences illustrate task-directed progress through repeated pickup and placement, complementing the fast dynamic behaviors described above. These tasks last over 40 seconds on average.
Frequently Asked Questions
What does Long-WAM use to guide robot actions? It conditions actions on observed history and anticipated futures, using prediction as an intermediate representation for control.
How does the context-scaling evaluation work? It varies the duration of observations in the causal prefix while keeping the visual forecast and action horizon fixed within each benchmark.
Which robots are used in the experiments? The experiments use a Unitree G1 humanoid and a YAM bimanual manipulator.
How does Long-WAM perform on Atomic-Seen tasks? Long-WAM alone achieves 67.9% Atomic-Seen success; querying it every 15 steps yields 84.4% without a planner.