Skill Library (ASPIRE-style)#

本节介绍 LeRobot 的通用功能。请按目标机器人、数据集和运行环境配置示例；用于 AlohaMini 2 / 2 Pro 前，需要确认硬件与接口兼容性。

Reusable robotic knowledge distilled from validated debugging sessions, in the format of ASPIRE (§2.2 Skill Library): each entry stores a failure signature (symptoms + trace evidence), a when-to-apply condition, a repair strategy, and a compact code sketch — not a full task program. Entries are meant to be retrieved as in-context guidance by a coding agent (or a human) when the matching failure signature shows up in execution traces.

Every skill here was discovered and validated on the AlohaMini Std/Pro robots in ManiSkill during scripted grasp/pick-place development (2026-07). The decisive diagnostics were per-primitive traces: IK residuals, PhysX contact-pair dumps (scene.get_contacts()), commanded-vs-actual joint errors, link world-poses, and rendered frames. This mirrors ASPIRE's robot execution engine feedback.

skill	one-line
`base_reposition.json`	target valid but out of the arm's envelope → drive the base to a feasibility-gated station (== ASPIRE's Multi-Angle Approach)
`station_physics_validation.json`	a station can be kinematically fine but physically untenable → teleport + settle-check before navigating
`tilted_approach.json`	vertical top-down descent bottoms the palm on the object → tilt the approach axis 55–70°
`desc_first_ik_branch.json`	grasp-pose IK seeded from the pre-grasp lands the wrist ON the object → solve the grasp config first, seed the pre-grasp from it
`cartesian_line_descent.json`	joint-space interpolation between IK waypoints sweeps an arc that rakes the object → IK waypoints along the approach line
`linear_push.json`	push a tabletop object to target XY without grasping -> feasibility-gated station, closed horizontal gripper, 1 cm Cartesian waypoints, contact/trajectory-tuned endpoint
`spurious_self_collision_bits.json`	convex-hull contacts between rigidly-related links shove the arm off command → share a collision-disable bit across the subtree
`empty_bbox_placement.json`	scene-placement Z computed from an empty robot bbox → robust bbox (extent union, instance proxies, ancestor walk)
`floor_penetration_base_pinning.json`	wheel collision geometry penetrates the floor and pins NAV with huge normal impulses → spawn the root at measured wheel-rest height
`station_relative_approach_dir.json`	tilted approach computed from a hardcoded base origin works only at home station → derive the approach from live arm-base pose
`elevated_release_place.json`	tilted-axis place drags the object during surface-contact descent → release 1.5–2.5 cm above the surface and let it drop
`frozen_arm_base_lift_place.json`	any arm reconfiguration while holding slings the object out → freeze the arm; fine-align with the base, descend with the lift

Loader: library.py (load_skills() returns all entries; match(symptoms) does a naive keyword match over failure signatures — good enough for in-context retrieval).