VLA Models (cross-module)

VLA (Vision-Language-Action) models unify vision, language, and action in one autoregressive generator. RT-2 discretizes each dimension of the 7-DOF end-effector pose (3D translation + 3D rotation + gripper) into 0–256 bin tokens appended to the vocabulary, generating action sequences jointly with image-text tokens; the quantization is: $$a_{\text{bin}} = \left\lfloor \frac{a - a_{\min}}{a_{\max} - a_{\min}} \times 256 \right\rfloor, \quad a_{\text{bin}} \in \{0, \dots, 255\}$$ OpenVLA fine-tunes a pretrained VLM to output discrete action tokens similarly. The alternative Diffusion Policy route uses a conditional diffusion model to generate continuous action trajectories directly, with training objective: $$\mathcal{L} = \mathbb{E}_{t, \epsilon} \left[ \lVert \epsilon_\theta(a_t, t, c) - \epsilon \rVert^2 \right]$$ denoising action trajectories conditioned on observations $c$.

robot manipulation and embodied-LLM interviews; often asked: 'how to represent continuous actions' and 'discretization vs diffusion'.

robot policy data is scarce, so VLA models directly reuse VLM visual-language priors; for action representation, discretization (0–256 bins) slots actions into off-the-shelf autoregressive architectures, while Diffusion Policy avoids discretization resolution loss through iterative denoising and is more robust to multimodal action distributions. See the RL module: world-model/RL loops in the RL module (guide model-based-rl-and-planning).

How does RT-2 discretize 7-DOF continuous actions into 0–256 bin tokens? Write the quantization formula?

Action discretization vs Diffusion Policy: essential difference and what each solves?

OpenVLA vs RT-2: differences (backbone, training data, action output)?

Why can VLAs reuse VLM priors? How are action and text tokens unified in training?

Write Diffusion Policy's diffusion training objective; its relation to DDPM image denoising?