A quadrotor policy learned to thread a moving narrow gap purely through differentiable simulation reached 99.3 percent success in sim and 80 percent over 10 real flights
AspireAI community · arXiv (cs.RO)
Flying a quadrotor through a narrow gap that is moving demands centimetre-level alignment and exact timing, and a success that depends on one narrow band of states is hard to learn from a sparse reward. This work rewrites the objective instead. A differentiable moving-gap formulation gates the centering, alignment and collision terms by signed distance to a time-varying gap plane, turning a sparse traversal event into a dense pathwise loss that gradients can flow through.
The resulting policy reads onboard visual and inertial measurements directly and outputs collective thrust and body-rate commands. It is trained in one stage from random initialization, with no teacher, no planner and no distillation. In simulation, pathwise gradients beat score-function gradients for this precision-critical task, giving a 99.3 percent success rate after a few minutes of training on a single GPU. The same policy then transfers to a real quadrotor without retraining, crossing a dynamic narrow gap from varied initial conditions using onboard sensing only, at an 80 percent success rate over 10 physical trials.
The honest limitation is the sample size: the entire real-world result is 10 trials, so 80 percent carries a wide interval and the paper does not report the simulation-to-reality gap as a controlled ablation. 'Transferred zero-shot' means the policy saw no real-world data, not that the gap geometry was guaranteed to match. The few-minutes-on-one-GPU claim is a simulation figure, and the differentiable formulation adds real cost to the training path.
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.