Refine and sync colliders for all variants

[T-K-233]

Summary

Refits the colliders of every Lite variant.

• Shoe (lite_biped, lite, lite_pro): seven r8 capsules across each sole, ends 4 mm inside the outline. This replaces the 11 spheres of the biped and the 2 boxes of lite and lite_pro. The optional flat-bottomed sole hull is in a separate PR, Ship an optional flat-bottomed sole hull for lite_biped #24.
• Body (all six variants): one capsule per segment, the same in every variant that has it.
• Excludes: physics.json takes an exclude list of body pairs. The four legged variants exclude pelvis ↔ <side>_hip_roll, because the hip-motor capsule overlaps both butt capsules at rest.
• Fresh Onshape export of lite, lite_pro, lite_bimanual and lite_biped_debug (see below).

Segment	Collider	Body
pelvis	r69 capsule over the pelvis block; r52 capsule across both hip-pitch motors, on their axes	pelvis
butt	r51 capsule flush with the rear face of the shell (24°)	<side>_hip_roll
thigh	r55 capsule, hip-yaw axis to knee axis	<side>_hip_yaw
shank	r44 capsule, knee to ankle	<side>_knee_pitch
torso	r74 and r58 spheres on the torso axis	chest
head	r97 capsule, 234 mm	head
upper arm, forearm	r35, r32 capsules (wrist_yaw has none)	<side>_shoulder_yaw, <side>_elbow_pitch
hand	2–3 capsules over the palm and the fingers	<side>_hand
shoe	7 r8 capsules across the sole (debug foot: the box)	<side>_foot

Fit rules: first fit a contact face flush (the sole, the rear of the butt). Then fit for the coverage of the whole shell. A capsule that crosses a joint is on the joint axis. Where two capsules touched but their shells did not, in rest poses and single-joint sweeps, the capsule ends became shorter by 10 mm or less. The radii did not change. Contacts where the shells also touch stay, for example the forearm on the hip motors with the arms down. Each scad file tells why its collider has its shape.

Impact

Variant	Colliders	Shell > 5 mm outside every collider (zero pose)
lite_biped	26 → 22	72% → 28%
lite_biped_debug	6 → 10	66% → 24%
lite, lite_pro	17 → 35	58–59% → 26–27%
lite_dummy	12 → 11	54% → 52% (the chest mount has no collider)
lite_bimanual	11 → 10	67% → 58% (the chest mount has no collider)

Shoe, measured on lite_biped against the 11 spheres. Drops are of a free 8 kg foot, 256 per case:

	11 spheres	7 capsules
underside within 3 mm of a collider	5%	36% (before the 4 mm inset)
step time, flat / rough terrain (mujoco-warp, 4096 worlds)	–	+3 to +5% / −23 to −25%
stays on a step edge (side over the edge, centre over the step)	58–60%	97%
lands turned over on ±2 cm bumps	12–14%	23–24%

The 4 mm inset of the capsule ends increased the turn-overs on bumps from 16–18% to 23–24%.

Names change for consumers. The thigh moves from <side>_hip_roll_collision_0 (old: a 250 mm capsule along the hip-yaw axis) to <side>_hip_yaw_collision_0. <side>_hip_roll_collision_0 is now the butt. lite-motion-tracking's LEG_COLLISION_EXPR must add hip_yaw, so that the thigh continues to collide. berkeley-humanoids/lite-motion-tracking#6 does this.

With the arm colliders on, a retargeted LAFAN1 clip puts the arms into the torso and pelvis colliders in more frames than before, mostly where the shells also touch. lite-motion-tracking collides the legs only.

Fresh Onshape export

The fresh export was converted offline (onshape-to-robot --convert) from its robot.pkl and part meshes, with the scad files of this branch. Changes against the previous export:

	Change
lite, lite_pro mass	They carry the lite_biped legs: the same link masses. Total mass 32.2 → 25.2 kg (lite) and 33.3 → 26.3 kg (lite_pro). Per leg: pelvis 3.97 → 2.10, hip_pitch 0.62 → 0.18, hip_roll 1.30 → 0.93, hip_yaw 2.09 → 1.54, knee_pitch 0.44 → 0.23, ankle_yaw 1.43 → 1.03, ankle_pitch 0.18 → 0.34, foot 1.21 → 0.44 kg.
lite, lite_pro shoe	The older shoe, with a 12 mm shorter heel, is gone. The foot mesh is now the biped shoe, vertex for vertex. Its part frame moved 5 mm, so the foot scad files are now those of lite_biped. The sole capsules do not move. Now they agree with the shoe outline.
Joint limits	ankle_yaw ±45° → ±90° (lite, lite_pro, as on the biped). shoulder_yaw ±90° → ±120°, and wrist_yaw ±120° on both sides, where one side stopped at 90° (lite, lite_pro, lite_bimanual). The lite_bimanual ros2_control limits follow.
lite_biped_debug	The foot is now a front and a rear part. Foot mass 0.326 → 0.310 kg, mesh within 4 mm.
lite_bimanual	Hand mass 0.1675 → 0.1677 kg.
Other	Joint origins and axes do not change. The other meshes change by 1–2 mm (a new simplification), with the same area.

lite_biped and lite_dummy are not re-exported. The "Sim Biped" configuration now gives the debug foot, so it does not give the lite_biped shoe. The other lite_biped links agree with the fresh export.

Check of the colliders. The export gives the exact part frame of every scad file. The frames that were inferred for <side>_hip_yaw, waist and <side>_elbow_pitch are correct to 2 µm. Against the assumed frame, the head frame turns 180° about the head axis. Thus an export put the head capsule 19.5 mm too far back. head.scad is now written in the exact frame (also in lite_dummy, which has the same Head part). With this fix, the export reproduces the colliders of this branch in all four variants.

How the colliders were written

The capsules were fitted in link frames. The scad files hold them in Onshape part frames.

• Existing scad files: the part frame of each of the 49 files was recovered from its URDF collisions, all within 7 µm.
• New scad files (<side>_hip_yaw; <side>_elbow_pitch; waist on the bipeds): they take the part frame of their neighbours.
  • The leg parts share one frame. The cached Onshape meshes of Hip Yaw and Waist confirm it to 0.3 mm.
  • The elbow takes the frame that the shoulder and wrist parts share.
• URDF hub: its collisions were computed from the scad files through the parser of onshape-to-robot, as an export does. Then the MJCF and xacro were regenerated.
• Consistency: the same part gives a byte-identical scad file in every variant.

This step did not use an Onshape export.

Test plan

• pytest: 1605 passed, 36 skipped (adds exclude tests)
• STE and clean-code pass: each scad file parses to the same shapes
• lite-description-generate <variant> --only urdf,mjcf,xacro reproduces the committed files for all six variants
• Every MJCF collider matches the fitted set to 7 µm and 0.006°
• Every existing scad file reproduces its old URDF collisions under the recovered part frames (49 files, ≤ 7 µm)
• Fresh Onshape export (lite, lite_pro, lite_bimanual, lite_biped_debug): the colliders do not change, after the head frame fix
• pytest after the re-export: 1605 passed, 36 skipped
• Confirm the wider shoulder_yaw and wrist_yaw limits (±120°) on the bimanual hardware before a deploy

🤖 Generated with Claude Code

https://claude.ai/code/session_01HpFUBhqSfbDuibPRVhscr9