REAL-TIME ROBOT EXECUTION WITH MASKED ACTION CHUNKING
Training
Some notations:
- $A$ (
action_curr): The ground-truth expert action chunk directly from the dataset. - $\tilde{A}$ (
action_pred_curr): The action chunk generated completely from scratch by the frozen, pretrained backbone (without LoRA). - $\hat{A}$ (
action_init): The mixed action chunk used for the “Self-conditioned Curriculum” $\gamma \sim \mathrm{Bernoulli}(\sigma),\ \sigma \in [0,1].\ \hat{\mathbf{A}}_t = \gamma \mathbf{A}_t + \text{stop-gradient}\left((1-\gamma)\tilde{\mathbf{A}}_t\right)$. - $u$ (
u_t_curr): The ground-truth flow target. Mathematically, it is simply the vector pointing from the sampled noise to the target action. - $\hat{u}$ (
pred_curr): The flow predicted by the new, LoRA-adapted policy ($\hat{v}_\pi$). - $\tilde{u}$ (
with_backbone): The flow predicted by the frozen, pretrained backbone ($v_\pi$).
我个人认为这里的loss完全多余因为$\tilde{\mathbf{u}}\tau$会被约掉$\mathcal{L}\Delta = \sum_d \frac{ \sum_{\tau=0}^{P-1} \left| m_d^\tau ( \mathbf{u}\tau - \tilde{\mathbf{u}}\tau ) - m_d^\tau ( \hat{\mathbf{u}}\tau - \tilde{\mathbf{u}}\tau ) \right|2^2 }{ \max\left( 1, \sum{\tau=0}^{P-1} m_d^\tau \right) }$。
一个值得学习的点在于这个delay d会被
Inference
$\mathbf{A}{t}^{\tau + \frac{1}{n}} = \mathbf{m} \odot \left( \mathbf{A}{t}^{\tau} + \frac{1}{n} \hat{\mathbf{V}}{\pi} \left( \mathbf{A}{t}^{\tau}, \mathbf{O}{t}, \tau \right) \right) + (1 - \mathbf{m}) \odot \mathbf{A}{t}^{p}$.