# yam_sim tasks

## Bootstrap (done 2026-07-02)

- [x] Read `agents/yam_sim/ROLE.md` + `agents/yam_sim/inbox.md`
- [x] Read `shared/GUIDELINES.md` (mostly known from inherited context)
- [x] Set status to `working`
- [x] Mine inherited memory → write `vault/para/yam/simulation.md`
- [x] Ping yams's inbox with FYI
- [x] Set status to `idle`

## Standing objective (2026-07-02)

Build a **MuJoCo YAM (single-arm) pick-and-place sim** that supports:
1. Sim evals of trained policies
2. Sim2real experiments
3. Auto-generation of pick-and-place training trajectories

Scene = YAM arm + one pick object (real-scanned cup/mug/etc. when available) + one place target (plate/box). Task = pick and place.

## Near-term milestones (draft)

- [x] **Assets sourced** (2026-07-02): YCB banana (`011_banana`) + YCB plate (`029_plate`), photogrammetry-textured, real-scaled, in `/data/cameron/para/robot/yam/sim_assets/objects/ycb/`. Anzu / TRI clone deferred — YCB is auth-free + robotics-standard so no reason to fight SSO.
- [ ] Phase 1 — **MVP scene** (~1 session): single YAM arm + YCB banana + YCB plate as targets, waypoint-based procedural trajectory + magnet-style attach; verify one rollout visually. Gripper = default parallel-jaw (finray postponed). IK via Mink (`lib/ik.py` pattern, site `grasp_site`).
- [ ] Phase 2 — **randomization + record**: randomize object/target pose, record N episodes into a `record.py`-compatible schema
- [ ] Phase 3 — **more objects**: swap in additional YCB items (mug 025, bowl 024, cracker box 003) once Phase 2 is solid
- [ ] Phase 4 — **sim-trained policy eval**: run our v4 model against sim rollouts, first sim2real numbers

## Deferred / on hold

- Anzu banana + tumbler imports (task #216 in the inherited list) — deferred; YCB covers the banana. Tumbler needs SSO-clone from Mac if we revisit; not blocking Phase 1.