single mention · model architecture
Continuous DiT-style action expert
Adds a continuous diffusion-transformer-style action expert to a robot-aware backbone.
The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See multimodal architecture for the methods filed beside it.
Filed under model architecture › multimodal architecture
Related methods
Used for
Methods the sources say this one serves — as a part, a setting, a variant or an enabler.
post-training attaches a continuous DiT-style action expert to this already robot-aware backbone
Supported by
Methods the sources say are used for this one.
- Autoregressive depth-token prefix conditioningconfigures it
adds an autoregressive depth-token prefix before action prediction
Each block contains bidirectional self-attention over the action chunk, cross-attention to the VLM context, and an MLP
- Flow matching for continuous action generationimplements it
learns a flow-matching velocity field over continuous robot action chunks
- Masking action tokens from expert conditioningconfigures it
The discrete action-token span is masked from the expert conditioning path
and 5 more
- Near-zero action-expert output initializationconfigures it
The final projection is initialized to produce near-zero velocity predictions at the beginning of post-training
- Noncausal transformer action experta form of it
The action expert is a noncausal transformer over a short action chunk
- Per-layer key-value conditioningpart of it
The action expert receives VLM context through per-layer KV conditioning
- Rotary Position Embeddingpart of it
The expert uses rotary position embeddings for the action sequence, which gives the denoising transformer an explicit ordering
- Time-conditioned DiT modulationpart of it
used to produce DiT-style shift, scale, and gate parameters for all three residual branches
Filed alongside
Other methods under multimodal architecture.