Roadmap (all entries are plans)
This page collects the plan: what the bench intends to deliver by which date, and the yardsticks success is measured against.
Plans are not results — the measured values live on the other pages and are joined in here only as a status column.
All dates and targets on this page are plans; measured values live on the other pages. Every entry below carries a PLAN badge.
Goals (G1–G6)
| Goal | Target (plan) | Measured today |
|---|---|---|
| G1 scale | PLAN ≥ 100 high-quality surgical procedure/task envs, each past the env gate and citing proclib; ≥ 1,000 env variants (task × patient anatomy × arm count × port layout × difficulty). | 170 registered, 53 green, 0 counted — registry |
| G2 quality | PLAN An env counts only at realism ≥ 8/10 with real anatomy and cited/calibrated physics; below the bar it stays gated-not-counted with its gap list. Every env passes the env gate (next section). | best judged realism 5/10; bar: green env AND realism receipt glm_score >= 8 |
| G3 multi-robot | PLAN ≥ 15 tasks that need 2–8 arms to cooperate. | 23 tasks for 2–8 arms (receipt R10); 0/23 counted toward G1 yet (receipt R10) |
| G4 agent-driven | PLAN (a) an agent env factory generates, gates and registers new envs with no human step; (b) agent baselines (LLM planner, VLA, ACT/scripted) evaluated on the whole bench with receipts. | 24 envs registered from the agent env factory source (factory-v1) |
| G5 paper package | PLAN Method, bench tables, figures, a real-output video and reproducible run commands by 12-04; submission and any public release is an owner gate. | paper page; targets on this page are plan dates |
| G6 beyond human | PLAN Every task family carries a cited human-expert baseline; tables report the agent / human-expert ratio; "beyond human" only where measured. | measured for 7 families (receipt R9) |
Weekly plan to the RSS 2026-12-04 package
| Week | Dates (plan) | Planned output | Measured today |
|---|---|---|---|
| W1 | 11-06..11-12 | PLAN env registry + unified API on i4h/Arena; inventory and port every source; env gate script; ≥ 30 envs green | registry live: 170 registered, 53 green, 44 red (see environments); receipt R12 records the W1 snapshot (63 registered / 10 green at write time) before the factory builds landed |
| W2 | 11-13..11-19 | PLAN multi-robot tasks (≥ 15); variant generator (patients, arms, ports); agent env factory v1 builds proclib to-build items; ≥ 60 envs, ≥ 500 variants | 24 envs registered from the factory source; multi-arm task set measured in R10 |
| W3 | 11-20..11-26 | PLAN agent env factory at scale; best-backend realism upgrades (blood/fluid, tissue, render); baselines on NRP; ≥ 100 envs, ≥ 1,000 variants | (not started - plan) |
| W4 | 11-27..12-04 | PLAN freeze 12-01; tables, figures, video, paper draft; reproducibility pass; owner gate for submission | (not started - plan) |
Hard target dates and yardsticks
| Id | Target (plan) |
|---|---|
| G1-D-S30 | PLAN >= 1,000 COUNTED envs by 2026-11-30 (paper number freeze); >= 10,000 for the next version (RSS final 2027-04-16 / bench v2) |
| G2 | PLAN an env counts only at realism >= 8/10 with real anatomy and cited/calibrated physics |
| G3 | PLAN >= 15 tasks needing 2-8 cooperating arms |
| RSS-stage1 | PLAN RSS 2026-12-04 Stage 1 extended abstract due (6 pp = 5 + 1 refs, verified CFP) |
Plan sources
- BENCH.md — branch contrib-flow-20261106 (commit 646ef76bb092…, blob 71c8b7d38f4b…).
- Plan document — /home/robot/workspace/agent-next/task-runs/20261104-surgery-multiarm/BENCH-PLAN-RSS-20261204.md (sha256 7d54d12d95c8).