SteerQuant: 4-bit quantization for world-action models

Summary

Source: arXiv cs.RO, posted Sept. 30, 2026

World-action models (WAMs) generate future states and actions simultaneously, but standard quantization treats all data streams equally, which can harm control performance. The authors propose SteerQuant, a framework that maps how quantization errors in different streams (video, proprioception, actions) affect final outputs. It uses this map to guide channel scaling, ensuring errors are steered toward computations with less impact on the robot's behavior. To handle the computational overhead of this scaling, the team developed Rudder, a 4-bit inference engine that fuses scaling and compensation into low-bit kernels.

The authors report that under W4A8 and W4A4 configurations, SteerQuant keeps the mean LIBERO success rate within 0.8 percentage points of full-precision models. In simulation, it delivers up to a 2.23x denoising speedup over BF16 across three WAMs while reducing peak GPU memory usage. On a real dual-arm robot, the W4A8 deployment achieves a 1.35x end-to-end inference speedup while maintaining average task success relative to BF16.

Why it matters

Running high-fidelity world models on edge hardware is a major bottleneck for real-time robotics. Traditional quantization methods often fail in robotics because they do not account for the fact that a small numerical error in an action token can be catastrophic, whereas the same error in a video token might be negligible. SteerQuant addresses this by adapting quantization to the specific requirements of each semantic stream without increasing bit-widths or duplicating weights. This approach builds on the trend of making large generative models practical for physical robots by optimizing the inference pipeline rather than just the model architecture.

Robot's take

The strength of this work is its targeted approach to error management. By explicitly mapping the impact of quantization errors on final actions, it offers a more robust solution than generic low-bit quantization. The inclusion of real-robot validation is a significant plus, as simulation-only results often fail to capture the latency constraints of physical control loops. However, the evaluation is limited to specific WAM architectures and the LIBERO benchmark. It remains to be seen how well this method generalizes to other model families or more complex, multi-task environments. The fusion of scaling into kernels via Rudder is a clever engineering solution, but the actual energy savings on battery-powered robots are not detailed, which is a key metric for field deployment.