Action Discretization And Token Based Policy Representation

via “action-as-text-token-representation”

Google's vision-language-action model for robotics.

Unique: Represents robot actions as discrete tokens in the language model vocabulary rather than using continuous outputs or separate policy heads, enabling the same transformer decoder to generate both language and actions

vs others: Simplifies architecture compared to models with separate policy networks or continuous action heads, enabling more efficient joint training on language and robotic tasks within a single transformer framework

via “policy-constrained transaction execution with approval workflows”

Give your AI agent a wallet. AgentFi provides 10 MCP tools for executing DeFi transactions on EVM chains (Ethereum, Base, Arbitrum, Polygon). Swap tokens, transfer assets, supply to Aave, check balances and prices — all policy-constrained and simulated before broadcast. Each agent gets a dedicated S

Unique: Implements server-side policy rule engine that validates transactions against agent-specific schemas before Safe wallet execution, enabling fine-grained spending controls and approval workflows. Most agent frameworks lack built-in policy enforcement; developers must implement custom guards.

vs others: More flexible than fixed spending limits because policies can encode complex rules (token whitelists, counterparty restrictions), while faster than human-in-the-loop approval for low-risk transactions due to automatic approval for policy-compliant actions.

via “action discretization and token-based policy representation”

## Historical Papers <a name="history"></a>

Unique: Uses 8-bit discretized action tokens instead of continuous action regression, treating action generation as a categorical prediction problem. This leverages the transformer's native strength in discrete sequence modeling and enables efficient beam search or sampling-based action selection.

vs others: More sample-efficient and stable than continuous action regression in transformers, and enables efficient multi-hypothesis planning via beam search, though at the cost of quantization error and reduced precision compared to continuous approaches.