Action Primitives and Tools#

The planner operates through tool calls. Action primitives perform motions such as grasping, moving an end effector, or opening a gripper. Perception and state tools locate targets, read images, and inspect results.

Choose Action Tools#

RPent provides two main types of action tools:

  • VLA actions use a vision-language-action model to generate action sequences, such as LIBERO’s pi0_pick. The model usually runs in a separate service.

  • Scripted actions execute parameterized motions, such as move_to, rotate_wrist, and release. Names and arguments depend on the environment.

back_project, segment, and view_env_state are perception or state-reading tools; they do not themselves represent robot motion. finish ends the planner loop. Task success is determined by the environment or the real-robot operator.

Models by Platform#

Environment / robot

Action model and configuration

LIBERO

Pi0.5

RoboCasa365

RLDX-1

RoboTwin

LingBot-VLA

Single-Arm Franka

Pi0.5 / openpi

Dual-Arm Franka

Pi0.5 / openpi

Each platform page provides checkpoint paths, available tools, and task requirements. YAM deployment instructions and SO-101 content are pending; see YAM and SO-101.

Services and Extensions#

VLA services expose action prediction through predict and health checks through healthz. RPent RPC services support HTTP and socket transports; see Remote Services and Parallel Runs for deployment.

To add a tool, define its arguments, execution, and result handling as described in Add an Action Primitive. Planners call these tools through the shared toolkit.