Enterprise AI Analysis

OAT is a learned autoencoder framework that discretizes action chunks into an ordered sequence of tokens, uniquely satisfying all three desiderata: high compression, total decodability, and a structured ordering over tokens. It uses transformer-based register tokens to aggregate temporal information, Finite Scalar Quantization (FSQ) for a discrete bottleneck, and nested dropout along with causal attention to explicitly induce ordering. This ordered latent representation ensures that earlier tokens capture coarse, global structure, while later tokens refine finer details, enabling prefix-based decoding and an anytime trade-off between inference cost and action fidelity. The tokenizer (T) and detokenizer (T^-1) are jointly trained end-to-end using a reconstruction objective, making T^-1 a well-defined total function.

Experiments across 20+ simulation and real-world manipulation tasks demonstrate OAT's superior performance. Autoregressive policies equipped with OAT consistently outperform prior tokenization schemes (Binning, FAST, QueST) and diffusion-based baselines (DP). OAT exhibits a clear monotonic performance improvement as the number of decoded tokens increases, allowing for an anytime trade-off between computation and performance. Its ordered token space aligns with next-token prediction, facilitating stable and scalable autoregressive learning. OAT provides greater flexibility at inference time, enabling policies to generate valid actions even with partial token sequences, and progressively refine actions for higher fidelity.