Robot Foundation Models

OK-Robot’s modular pick-and-drop design makes the first failed handoff visible: retrieval, navigation, grasp proposal, execution or drop. Log the boundary before replacing a model.

HoMeR uses absolute 3D keyposes for long reaches and relative action deltas for fine contact. Label the phase handoff and distinguish policy errors from whole-body execution.

OpenVLA-OFT generates continuous action chunks in parallel. Check chunk timing, action units, inputs, compute and multimodal-task needs on the target robot before choosing it.

RMA estimates a latent adaptation signal from recent states and actions while keeping deployed policy weights fixed. Evaluate the transition after a terrain or payload change.

NoMaD’s goal mask switches a shared visual policy between undirected exploration and goal-image pursuit. The mission layer still selects subgoals and decides when to switch.

ViNT maps image goals, local coordinates and route commands into a navigation policy’s goal interface. Compare labels and errors by modality before using a new route source.

GNM-style cross-robot navigation needs aligned camera history, action scale and robot context. Audit each trajectory before mixing data, then test transfer by platform.

For openpi remote inference, keep image preprocessing and state normalization on the sides expected by the served checkpoint, then verify action-chunk timing and units before robot use.

VoxPoser turns instructions into 3D planner objectives. Trace grounding, value maps, waypoints and control, and identify what the released simulator does not supply.

SayCan combines a skill’s relevance to an instruction with its feasibility in the current state. Learn what those scores mean and why execution still needs verification.