feat(sim): add deformable sock prop - #729
Conversation
Greptile SummaryThe PR adds a deformable MuJoCo sock prop and its simulator lifecycle, browser skinning and streaming pipeline, packaged assets, and delivery validation.
Confidence Score: 4/5The PR is not yet safe to merge because the sock remains temporally ahead of the gripper during delayed playback. The rigid robot and prop state is rendered on a delayed simulation timeline, but deformable updates still retain and apply only the newest frame, so motion while grasping the sock can produce visible sliding or finger penetration. Files Needing Attention: sim/viewer/src/simSession.ts Important Files Changed
Reviews (3): Last reviewed commit: "fix(sim): localize soft sock deformation..." | Re-trigger Greptile |
| for (const id of this.#deformablesDirty) { | ||
| const frame = this.#deformables.get(id); | ||
| if (frame) scene.setDeformableFrame(frame); | ||
| } | ||
| this.#deformablesDirty.clear(); |
There was a problem hiding this comment.
Deformable playback bypasses timeline
When the robot moves while grasping the sock, these lines apply the latest deformable frame after rendering the robot on the delayed playback timeline. This places the sock 25–250 ms ahead of the gripper, causing visible sliding or finger penetration during motion; buffer and select deformable frames by their included simulation timestamp before applying them.
Knowledge Base Used: Simulation viewer and world state
d8ec95e to
307a8c0
Compare
307a8c0 to
ead1468
Compare
Summary
Validation
./innate-sim up.sim/assets.lock.json.ci/seed_asset_context.py; Publish Simulator Images and Publish Viewer Bundle are green, and the geometry image is available in GHCRHandoff: please try this grasp
Place Soft sock, pinch it around the heel/body with both gripper fingers, then move the arm while holding it. Please check:
Measured locally, free/drop simulation had about 2.1x real-time capacity. A heavy active pinch ran around 0.66x real time. Denser physics cages were tested and rejected because they reduced performance as far as 0.22x real time.