# PARA — goals + narrative

## Long-term vision

Cameron's two-year arc: ship PARA as a banger paper → use credibility + TRI relationships → first signs of life for a household robot company by mid-2027. Identity-level frame (per life_manager 2026-05-25): become the version of yourself you'd hire at 27.

PARA is the technical lever. The startup is the destination. The internship at TRI (2026-05-26 → 2026-08-17) is where the lever gets tested and the relationships get built.

## Paper narrative (locked framing)

Full pitch at `/data/cameron/para/pitch.md` (30-sec + 2-min versions, with anti-FAQ). Lab presentation deck at `omidlab.net/para_presentation` (22 slides, locked 2026-05-12).

**Opening hook** — pick based on audience:
- ML researcher: *"Modern image features generalize across viewpoints and positions. The policies built on them don't. We fix the gap."*
- Investor / non-technical: *"Our robot learns from 15-30 demonstrations instead of 3,000, because we predict actions on the image instead of regressing 3D coordinates."*

**Binding framing (canonical 2026-06-10)** — second-beat mechanism statement, subsumes the old factorization framing:

> *"When you train a robot policy from image-to-action labels alone, you ask it to jointly learn — purely from those labels — the EEF↔pixel association, the depth along that pixel's ray, the 3D scene geometry, AND the implicit camera-to-robot extrinsics. PARA **binds the action space to its projecting image features**: discretizes the volume of candidate EEF positions, attaches each candidate to the pretrained image features at its projection, and reformulates policy learning as feature **selection** over that volume. The geometry-learning burden disappears — one line of camera math hands it to the policy for free. Result: large data-efficiency and generalization gains, no depth sensor, no 3D backbone, dense-in-time 6-DoF trajectories. Price of admission: known robot-to-camera extrinsics."*

Use this after the paradox / data-efficiency hook, before the why-it-works list. Never use it as the opener (mechanism-first anti-pattern — see `vault/fleet/agents/project_highlevel/memory.md`).

**Why this re-framing**: subsumes both old PARA (factorized single-view binding via uv heatmap + height bin along one ray) and new PARA (joint multi-view binding via world voxel grid projected into every view) as instances of the same family — *action selection on a feature-bound candidate volume*. View count becomes an implementation detail of the binding, not a story. Dissolves the Jun-3 single-view headline tension cleanly.

**Pitch-anchor phrase**: *"bind the action space to its projecting image features"* is the load-bearing sentence. The "selection vs regression" distinction is what makes the data-efficiency claim mechanistic instead of magical.

**OLD — Second-beat factorization framing (canonical 2026-06-02 → 2026-06-09)** — superseded by binding framing above. Preserved for reference:

> *"PARA is a new robot action head. We factorize the target EEF position (x, y, z) into (a) the pixel uv predicted as a heatmap over the image, and (b) the world-frame height predicted as a categorical distribution over discretized height bins along that pixel's ray. The single critical choice is height (world-frame Z), not depth (camera-frame distance) — that's what bakes in viewpoint invariance."*

The old framing is now the special case for single-view PARA. Height-vs-depth is still true but no longer load-bearing for the headline — it's a property of how the binding's height axis is constructed in world frame.

**Three promises that fall out of the architecture choice** (in order of paper section):
1. **Spatial OOD generalization** — robust to held-out object positions, camera viewpoints, environments.
2. **Video models become policies** — pixel-aligned head decodes actions from any model that predicts pixels (SVD, etc.). 90% vs 0% on same features.
3. **Cross-embodiment via pixel-aligned supervision** — supervise from non-robot data (arm-deleted point tracks: 42 vs 0). *Note: UMI exploration was dropped 2026-05-11 after generalization failed. Bench this contribution unless it resurfaces empirically.*

**Strategic framing (added 2026-05-25)**: lead with **data efficiency** as the headline. Sticky number target: *"15 demos vs 3,000+."* The 12-week TRI plan optimizes for landing this single sentence on slide 1.

**Inductive-bias framing (added 2026-05-19, from lab deck slide 2)**: *"Did we skip a step?"* — DUSt3R-level inductive bias in the output representation; PARA picks pixel-aligned actions instead of pointmaps. Composes with the head-vs-backbone result.

## New paper claim — "head > backbone" (added 2026-05-19)

Smaller DINOv3-S/16+ (30M) with KV-factored head **beats** DINOv3-L (305M) with dense head — 8.70 / 9.32 px vs 9.88 px train. This is a *new* paper claim worth a paragraph in the abstract, not just the ablations.

## OLD — Single-view headline decision (canonical 2026-06-03 → 2026-06-09)

> **Retired 2026-06-10.** The binding framing makes view count an implementation detail of how candidates get their bound features, not a headline claim. Today's canonical arch consumes 2 views (scene + wrist) and the pitch no longer takes a position on view count. The "humans use ego view" rhetorical move is dropped; "PixelNeRF was always multi-view" replaces it. Preserved below for reference.

**Wrist cam dropped from the headline.** Single third-person scene cam is the paper's headline modality. Two reasons:
1. **Engineering reality**: wrist cam live deploy was finnicky vs scene cam in TRI Wk 1 testing.
2. **Rhetorical move**: "humans use ego view, not wrist cam — single view isn't theoretically impossible." Pre-empts the "but you should fuse wrist cam!" reviewer comment. PARA isn't *weaker* than fusion methods — it's *proving fusion isn't necessary*.

This frees Phase 2 of the TRI plan (was wrist-cam integration). Time reallocated to single-view depth (more OOD axes, more dexterous tasks, more backbones). Wrist cam may resurface as ablation / appendix only.

## Locked overview figure structure (2026-06-03)

Top half: **PARA action head vs global regression head** diagram. This IS the second-beat factorization framing made visual.

Bottom half: **4-5 evidence buckets**, each with rollout stills + numbers strip:
1. **Data efficiency** — sticky number ("15 demos vs 3,000+")
2. **OOD object position** — train left → test right, workspace-split viz + PARA/ACT rollout stills
3. **OOD viewpoint** — train one cam pose → test held-out cam pose, side-by-side stills
4. **Dexterity** — 2 tasks max (cap): towel fold (SO-100 anchor 97/11) + cup/mug (YAM). Egg-carton / desk-cleanup / bimanual are *aspirational*; don't put in figure until results land.
5. **Flexible backbones** *(optional 5th bucket)* — same head on DINOv2 / DINOv3 / VGGT (VGGT aspirational). Composes with head>backbone claim.

**Staging by evidence-readiness**:
- **V1 (this week)**: action head diagram + buckets 1 (real) + 2-3 (placeholders pending YAM eval) + 4 (towel fold anchor only)
- **V2 (end TRI Wk 2)**: buckets 2-3 with real YAM evidence
- **V3 (paper submission)**: bucket 4 with desk cleanup added IF Phase 4 succeeds; bucket 5 with VGGT IF it lands

**Scope discipline**: 2-task dexterity bucket cap. Don't let the figure become a wishlist — "2 dexterous tasks + clean numbers" reads as evidence; "5 tasks half-supported" reads as scrambling.

## FidExoskeleton paper V2 (Week 11)

Side deliverable scheduled in `tri_internship_plan.md`. ~1 day, scope-capped. Adds YAM board file + "PARA elevates camera-pose calibration from hygiene to load-bearing" framing. Don't let scope expand.

## Paper venue + deadline

TBD — pending Cameron's commitment. Probable target: CoRL 2026 or RSS 2027 depending on timing.

## See also

- `/data/cameron/para/pitch.md` — full pitch with FAQ
- `/data/cameron/para/notes/tri_internship_plan.md` — how the goals decompose week-by-week
- `omidlab.net/para_presentation` — current lab talk deck