SpaceMind

Fly around a visible target and station directly above it at a safe distance. Use for reposition tasks that must end overhead of the target rather than in front of it.

Always-on operational baseline for SpaceMind. Defines shared coordinates, action policy, safety guards, observability rules, and perception workflow for all tasks.

Image-first laboratory navigation policy for ground-robot experiments. Use when the task must be solved with raw current images, conservative motion, and no oracle geometry.

Recover a temporarily lost target in the laboratory with bounded planar search and no blind forward motion.

Safely approach a visible laboratory satellite target and stop in a conservative near-front hold using only current images and motion tools.

Diagnose the laboratory target from raw image evidence, conservative viewpoint changes, and optional code support.

Find the satellite with bounded image-only search and then approach it conservatively to a safe near-front stop.

Add ReAct-style reasoning/output behavior on top of the active task skills. Use when the run is in react mode and the model should think briefly in text while still preferring real function calling.

Plan with a small world-model search before acting. Use for world_model planner calls that must compare exactly three candidate actions and pick the best one in JSON.

Select a single best tool action when the world-model planner fallback is needed. Use for world_model selector calls that must output one tool action as JSON.

Guide the agent to effectively use execute_code() for computation-assisted decision making during navigation and inspection tasks.

Always-on operational baseline for SpaceMind. Defines shared coordinates, action policy, safety guards, observability rules, and perception workflow for all tasks.

Choose sensing cadence and translation magnitude based on estimated distance. Use for approach tasks when the model must avoid overly conservative fixed step sizes at long range.

Plan with a small world-model search before acting. Use for world_model planner calls that must compare exactly three candidate actions and pick the best one in JSON.

Select a single best tool action when the world-model planner fallback is needed. Use for world_model selector calls that must output one tool action as JSON.

Guide multimodal observation for spacecraft situational awareness. Use for perception-heavy tasks that combine image quality, segmentation, local inspection, and knowledge-base reasoning.

Recover the target when it is lost or poorly centered. Use for navigation tasks where the agent must regain visibility without wandering indefinitely.

Diagnose the target spacecraft using multimodal evidence. Use for inspection-diagnosis tasks.

Reach a stable front hold point near the target spacecraft. Use for visible-target approach tasks that must end at about 2m from the target surface.

Search for the target when it is initially outside the field of view, reacquire it safely, then approach to a stable front hold.