# PARA — overview

Pixel-Aligned Robot Actions. Predict EEF as a keypoint in image space using DINOv3-feature volumes; lift to 3D via height bins + camera geometry.

## Current state (2026-06-10)

- **Canonical architecture (TRI / YAM, since 2026-06-10)**: **multi-view volumetric model with per-voxel grip/rot heads**. 2 views (scene cam fixed via aruco PnP + right-wrist via FK + hand-eye), shared DINOv3 ViT-S/16+ backbone, world-space voxel grid (V ≈ 1M, N=2 view-slabs × Z=64 height bins × P=128² pixel grid). Cross-view fusion via voxel→view projection + bilinear sample of per-pixel feats. T=32 per-timestep softmax over voxels (CE on GT-nearest voxel). Grip + rot are now **per-voxel heads** gathered at the GT (train) / argmax (infer) voxel + 3-layer 1D CNN over T. Deploy uses action chunking depth=4 + closed-form reverse-projection volume pool (~5 ms/chunk). See `vault/feedback_multiview_voxel_canonical.md` (long-term memory). Replaced dino_kv production single-view arch.
- **Pre-TRI headline number (smith300, train-set 2D argmax pix err, OLD arch)**: **9.32 px** with sin/sin S/16+. Beats DINOv3-L + dense conv head (9.88 px) — **head > backbone** is a paper claim worth surfacing. Smith300 numbers stay anchored to the old arch.
- **In-progress at TRI (Week 3 of internship, 2026-06-10)**: 3 trainings queued — cup_pluswrist_moredata (50 ep, GPU 0), cup_pluswrist_25ep (25 ep, GPU 1, data-efficiency probe), towel_fold_basic_wrist (30 ep, auto-launches when first finishes). New canonical arch's first real test.
- **Open narrative question (raised 2026-06-10)**: Jun 3 locked single third-person scene cam as the headline modality; today's canonical arch consumes 2 views. Pitch needs re-frame — either "two views was always right, the contribution is the per-voxel head" or "wrist is supervisory" (false — deploy uses both). Unresolved; brainstorm pending.
- **Blocked on**: nothing critical. Right-arm `T_left_from_right` convention debug paused (yams owns; not load-bearing for current experiments).

## Active surfaces

| Slice | Canonical path | Owner agent |
|---|---|---|
| Model code (libero + smith300) | `/data/cameron/para/libero/` | backbones |
| Paper LaTeX | `/data/cameron/para/paper/` | paper_writer |
| Figures + media | `/data/cameron/para/.agents/reports/...media/` | figure_maker |
| Website | `/data/cameron/para/.agents/reports/project_site/` | website_builder |
| Lab presentation (22 slides, locked 2026-05-12) | `omidlab.net/para_presentation` | project_highlevel |
| Mac-side deploy (smith300, izzy, UMI) | `~/Projects/robotics_testing/para_volume_kv/` on Mac | mac |
| TRI / YAM deploy + experiments | `/home/robot-lab/cameron/...` | yams |

## Anchor results (existing, pre-TRI)

- Real robot SO-100 (20 demos): 97% / 9% / 95% PARA vs ACT on pick-place / fold-towel / wipe-table
- New environment never seen: 94% vs 0%
- Zero-shot new viewpoint: 52% vs 0%
- Video backbone adaptation: 90% vs 0% (PARA head vs global regression on same SVD features)
- LIBERO position OOD left→right: 54% vs 1%
- LIBERO viewpoint cliff chart: PARA plateau ~62% through 18°, ACT cliff to 0%

Anchor numbers live in fig1 right-panel until YAM cup-task numbers replace them post-TRI.

## See also

- `goals.md` — paper narrative, pitch, venue
- `tasks.md` — active workstreams
- `memory.md` — PARA-specific conventions (rotation discretization, image norm, headline metric, etc.)
- `/data/cameron/para/notes/tri_internship_plan.md` — 12-week TRI plan
- `/data/cameron/para/notes/arm_v2_plan.md` — post-paper custom arm rebuild