A simulator that generates event-camera frames directly from scene geometry runs 400 times faster than rendering, and the policy it trained fits on a 1.2 MB drone
AspireAI community · arXiv
*Event cameras are the sensor of choice for small fast robots, but simulating them means rendering photorealistic frames and then converting those to events - which is exactly the step that makes large-scale reinforcement learning impossible.* This analytic simulator skips the rendering entirely, synthesizing event streams directly from scene geometry and ego-motion on the premise that event cameras fundamentally measure moving contrast edges. It generates a 64x64 event observation in 17 microseconds on a single GPU, over 400 times faster than a state-of-the-art rendering pipeline, which brings goal-directed obstacle-avoidance RL convergence under 30 minutes - three orders of magnitude faster than training on rendered events. The policy transfers zero-shot to ray-traced event simulation, reaching 83 percent success in cluttered courses unseen during training at 2.4 m/s average goal velocity, and retains 70 percent in a texture-dense indoor hall. Quantized to a 1.2 MB int8 network it runs fully onboard a 56-gram Crazyflie with no fine-tuning or external processing. The limitation: the observation is a coarse 64x64 grid, and the transfer claim is to simulation rather than to physical hardware beyond the small drone.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.