# 2026-05-26 (Tue) — Day 1 at TRI

## Setup

- DGX access granted (head + compute nodes, docker group). Username `cameron.smith`, default password (rotate when prompted).
- YAM workstation access granted (`robot-lab` shared account).
- `tri/` workspace scaffolded on dev machine.
- `~/.ssh/config` updated with TRI host blocks (dgx, dgx01, dgx02, robot-lab).

## Today / tomorrow

- [ ] Smoke-test SSH to all four host aliases
- [ ] First Raiden visualizer launch on sample data
- [ ] First conversation with Sergey or other team members re: docker base image, GPU reservation etiquette, lab safety

## Notes

(free-form — what landed, what surprised, who you met)

### Calibration verification (yams agent, evening)

Took over from project_highlevel mid-day; v6 had `LEFT_EE` aligned in pickup_apple but `RIGHT_EE` apparently wrong, and the next task was to test v7 (invert `T_left_from_right`).

**Outcome: v6 was already correct; the apparent "wrong" was a false negative.**

- v7 inversion contradicts raiden's own visualizer (`visualizer.py:194,257,394` applies `T_left_from_right` forward, no inversion).
- pickup_apple's right arm is parked off-frame in `scene_1` → projected `RIGHT_EE` lands in empty workspace, looks wrong.
- Switched to `BlockOnBlockRightArmYAM/0000/scene_camera` (right-arm-active task) — v6 projection lands dead-on the right gripper, including at frame 600 where the gripper is actively clamping a blue block.
- Sanity checks: left EE on left gripper, right base origin on right base mount, all consistent.

Surprise: `T_left_from_right` translation differs *a lot* across recordings (~10cm in pickup_apple vs ~60cm in BlockOnBlock). It's re-calibrated per session. Don't hard-code it; always read from the lowdim pkl.

Surprise: camera names also differ across tasks (`scene_1` vs `scene_camera`, `left_wrist` vs `left_wrist_camera`). Always read from `metadata_shared.json`.

Updated `tri/processes/data_pipeline.md` with the action layout, the convention rule, and the verification protocol. Verification crops in `tri/journal/assets/v6_verify_*.png`.

### Tomorrow / next

1. Run `rd visualize --web` on pickup_apple as the ground-truth cross-check before relying on any custom overlay further.
2. Begin porting PARA deploy stack to YAM (joint state + control interface).
3. Lab tour / first conversation with Sergey.