Skip to content

Promote reference-runs v1.0 backfill to main #203

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
KraHsu merged 2 commits into from
Jun 1, 2026
Merged

Promote reference-runs v1.0 backfill to main #203

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
51c72dd
Backfill reference-runs docs with v1.0 numbers
KraHsu Jun 1, 2026
07ba7e2
Merge pull request #202 from KraHsu/feat/genesis-1.0-reference-numbers
KraHsu Jun 1, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
71 changes: 36 additions & 35 deletions docs/best-practices/reference-runs.en.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -109,32 +109,31 @@ The tables below are the **v1.0** Genesis numbers. The v0.4.7 numbers are
preserved in a "Previous: v0.4.7" admonition next to each task so a
re-baseline diff is one scroll away.

!!! note "Status: v1.0 sanity smoke, full sweep pending"

The v1.0 reference cells stay `TBD` until the full
30 000-iter / 4096-env sweep is rerun on Genesis 1.0. In the meantime
a **50-iter G1-Velocity-Flat sanity smoke** was run during the
migration (S1, PR #175) and shows a healthy learning trajectory:

- Mean reward: iter 36 **−1.08** → iter 49 **−0.79** (monotonic improvement)
- Throughput: **~53 000 steps / s** sustained on RTX 4090 (CUDA 13.0)
- Wall-clock: ~1 min 39 s for 50 iters

The sanity smoke confirms the v1.0 simulator runs the existing G1
policy / rewards / observations without regressions in the hot
path — the only thing left is the multi-day wall-clock to populate
the per-seed convergence numbers. Maintainers running the full
sweep should fill the `TBD` cells from the corresponding
`eval.json` and append a footer with the GPU + driver pair the
numbers were collected on.
!!! note "Hardware the v1.0 numbers were collected on"

- **Cartpole IP / DIP, Franka** — NVIDIA H200 (141 GB HBM3), driver
570.211.01, CUDA 12.8, `QD_GRAPH=0` (Hopper requires it — no
`graph_do_while` fatbin for SM 90). Cartpoles ran with
`--parallel 3` on a single GPU; Franka ran with one seed per GPU
across GPUs 0–2.
- **G1-Velocity-Flat, G1-Tracking-Flat** — 4× NVIDIA GeForce RTX
4090 (24 GB each), driver 580.159.03, CUDA 12.8. G1-Velocity used
one seed per GPU on GPUs 0–2; G1-Tracking ran seed 1 alongside
Velocity on GPU 3 (Phase A), then seeds 2/3 on GPUs 0/1 after
Velocity finished (Phase B). The cross-task concurrency on one
host oversubscribes the CPU per this doc's own warning — the
reward numbers are deterministic and unaffected, but the train
wall-clock below is the **concurrent-run** wall, not a solo
figure (a solo run on the same GPU is ~3–6× faster per the
`Time elapsed` counter in `train.log`).

### `GeneLab-Inverted-Pendulum-v0`

| Seed | Final `return_mean` | `return_std` | Convergence iter | Train wall-clock | Eval wall-clock |
|---|---|---|---|---|---|
| 1 | TBD | TBD | 150 | TBD | TBD |
| 2 | TBD | TBD | 150 | TBD | TBD |
| 3 | TBD | TBD | 150 | TBD | TBD |
| 1 | 39.977 | 0.002 | 150 | ~2.5 min | 12.6 s |
| 2 | 39.994 | 0.001 | 150 | ~2.5 min | 12.4 s |
| 3 | 39.986 | 0.001 | 150 | ~2.5 min | 12.6 s |

Eval `length_mean = 1000.0` for all seeds (episode hits the time-limit cap
without falling), so the policy is solved at the budget cap. `success_rate`
Expand All @@ -152,9 +151,9 @@ is `null` (task does not publish `extras["is_success"]`).

| Seed | Final `return_mean` | `return_std` | Convergence iter | Train wall-clock | Eval wall-clock |
|---|---|---|---|---|---|
| 1 | TBD | TBD | 300 | TBD | TBD |
| 2 | TBD | TBD | 300 | TBD | TBD |
| 3 | TBD | TBD | 300 | TBD | TBD |
| 1 | 59.933 | 0.020 | 300 | ~3.5 min | 16.3 s |
| 2 | 59.914 | 0.174 | 300 | ~3.5 min | 16.4 s |
| 3 | 59.897 | 0.023 | 300 | ~3.5 min | 16.4 s |

Eval `length_mean = 1200.0` for all seeds. `success_rate` is `null` (same
reason as IP).
Expand All @@ -171,9 +170,9 @@ reason as IP).

| Seed | Final `return_mean` | `return_std` | Convergence iter | Train wall-clock | Eval wall-clock |
|---|---|---|---|---|---|
| 1 | TBD | TBD | 30 000 | TBD | TBD |
| 2 | TBD | TBD | 30 000 | TBD | TBD |
| 3 | TBD | TBD | 30 000 | TBD | TBD |
| 1 | 112.038 | 4.816 | 30 000 | ~28 h | 163.5 s |
| 2 | 112.871 | 4.918 | 30 000 | ~28 h | 160.3 s |
| 3 | 113.163 | 4.850 | 30 000 | ~28 h | 160.9 s |

Eval `length_mean = 1000.0` for all seeds (play_env `episode_length_s =
20 s` × 50 Hz). `success_rate` is `null`.
Expand All @@ -190,9 +189,9 @@ Eval `length_mean = 1000.0` for all seeds (play_env `episode_length_s =

| Seed | Final `return_mean` | `return_std` | Convergence iter | Train wall-clock | Eval wall-clock |
|---|---|---|---|---|---|
| 1 | TBD | TBD | 30 000 | TBD | TBD |
| 2 | TBD | TBD | 30 000 | TBD | TBD |
| 3 | TBD | TBD | 30 000 | TBD | TBD |
| 1 | 138.444 | 0.006 | 30 000 | ~28 h | 227.3 s |
| 2 | 137.980 | 0.008 | 30 000 | ~29 h | 301.6 s |
| 3 | 138.060 | 0.004 | 30 000 | ~29 h | 236.9 s |

Eval `length_mean = 1500.0`. The tracking play_env normally sets
`episode_length_s = 1e9` for infinite viewer playback; `genelab eval`
Expand All @@ -213,13 +212,15 @@ is `null`.

| Seed | Final `return_mean` | `return_std` | `success_rate` | Convergence timestep | Train wall-clock | Eval wall-clock |
|---|---|---|---|---|---|---|
| 1 | TBD | TBD | TBD | 2 000 000 (budget cap) | TBD | TBD |
| 2 | TBD | TBD | TBD | 2 000 000 (budget cap) | TBD | TBD |
| 3 | TBD | TBD | TBD | 2 000 000 (budget cap) | TBD | TBD |
| 1 | −6.122 | 13.307 | 0.990 | 2 000 000 (budget cap) | ~67 min | 10.7 s |
| 2 | −7.260 | 14.695 | 0.990 | 2 000 000 (budget cap) | ~62 min | 10.4 s |
| 3 | −8.790 | 18.275 | 0.970 | 2 000 000 (budget cap) | ~64 min | 11.1 s |

Eval `length_mean = 100.0` (fixed episode length). `success_rate` reflects
the goal-reach termination from the manipulation task; per-seed means and
the cross-seed mean will be filled once the v1.0 runs land.
the goal-reach termination from the manipulation task. Cross-seed mean
`success_rate ≈ 0.983 ± 0.012` — meaningfully tighter than the v0.4.7
sweep, where one seed only reached 0.89 from end-effector drift on harder
goal poses. All three v1.0 seeds land in the same tight band.

!!! note "Previous: v0.4.7"

Expand Down
64 changes: 33 additions & 31 deletions docs/best-practices/reference-runs.zh.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,28 +96,28 @@ GPU vectorized 慢很多。
下表为 Genesis **v1.0** 的数字。v0.4.7 的旧数字保留在每个任务旁边的
"Previous: v0.4.7" admonition 里，方便 re-baseline 对比。

!!! note "状态：v1.0 sanity smoke 已过，完整 sweep 待跑"

v1.0 reference 单元仍为 `TBD`，要等完整的 30 000 iter × 4096 env
sweep 在 Genesis 1.0 上重跑后才能填。期间已在迁移过程中（S1，
PR #175）跑了一次 **50-iter G1-Velocity-Flat sanity smoke**，轨迹健康：

- Mean reward：iter 36 **−1.08** → iter 49 **−0.79**（单调改善）
- 吞吐量：**~53 000 steps / s**，RTX 4090（CUDA 13.0）下持续
- 墙钟：50 iter 共 ~1 min 39 s

Sanity smoke 证明 v1.0 simulator 跑现有 G1 的 policy / reward /
observation 在热路径上没有回归；剩下的只是多天的墙钟跑完，填进各
seed 的收敛数字。维护者跑完完整 sweep 后，请把 `TBD` 单元从对应
`eval.json` 里取出填入，并附上跑数所用的 GPU + 驱动组合。
!!! note "v1.0 数字所用硬件"

- **Cartpole IP / DIP、Franka** — NVIDIA H200（141 GB HBM3），驱动
570.211.01，CUDA 12.8，`QD_GRAPH=0`（Hopper SM 90 没有
`graph_do_while` fatbin，必须设）。Cartpoles 用 `--parallel 3` 在
单张 GPU 上并行；Franka 用一卡一 seed 跑在 GPU 0–2 上。
- **G1-Velocity-Flat、G1-Tracking-Flat** — 4× NVIDIA GeForce RTX 4090
（每张 24 GB），驱动 580.159.03，CUDA 12.8。G1-Velocity 用一卡一
seed 跑在 GPU 0–2 上；G1-Tracking 的 seed 1 在 Phase A 与
Velocity 并行跑在 GPU 3 上，seed 2/3 在 Velocity 跑完后才在 GPU
0/1 上跑（Phase B）。跨任务在同一台机上并发触发了文档自身警告的
CPU 过订阅 —— reward 数字是确定性的不受影响，但下面的训练墙钟是
**并发跑的墙钟**，不是单 job 数字（同卡单 job 跑按 `train.log`
里的 `Time elapsed` 计数约快 3–6×）。

### `GeneLab-Inverted-Pendulum-v0`

| Seed | 最终 `return_mean` | `return_std` | 收敛 iter | 训练 wall-clock | Eval wall-clock |
|---|---|---|---|---|---|
| 1 | TBD | TBD | 150 | TBD | TBD |
| 2 | TBD | TBD | 150 | TBD | TBD |
| 3 | TBD | TBD | 150 | TBD | TBD |
| 1 | 39.977 | 0.002 | 150 | ~2.5 min | 12.6 s |
| 2 | 39.994 | 0.001 | 150 | ~2.5 min | 12.4 s |
| 3 | 39.986 | 0.001 | 150 | ~2.5 min | 12.6 s |

三个 seed 的 eval `length_mean = 1000.0`（episode 跑满时间上限不倒），策略
在预算上限解掉。`success_rate` 为 `null`（任务未 publish
Expand All @@ -135,9 +135,9 @@ GPU vectorized 慢很多。

| Seed | 最终 `return_mean` | `return_std` | 收敛 iter | 训练 wall-clock | Eval wall-clock |
|---|---|---|---|---|---|
| 1 | TBD | TBD | 300 | TBD | TBD |
| 2 | TBD | TBD | 300 | TBD | TBD |
| 3 | TBD | TBD | 300 | TBD | TBD |
| 1 | 59.933 | 0.020 | 300 | ~3.5 min | 16.3 s |
| 2 | 59.914 | 0.174 | 300 | ~3.5 min | 16.4 s |
| 3 | 59.897 | 0.023 | 300 | ~3.5 min | 16.4 s |

三个 seed 的 eval `length_mean = 1200.0`。`success_rate` 为 `null`（同 IP）。

Expand All @@ -153,9 +153,9 @@ GPU vectorized 慢很多。

| Seed | 最终 `return_mean` | `return_std` | 收敛 iter | 训练 wall-clock | Eval wall-clock |
|---|---|---|---|---|---|
| 1 | TBD | TBD | 30 000 | TBD | TBD |
| 2 | TBD | TBD | 30 000 | TBD | TBD |
| 3 | TBD | TBD | 30 000 | TBD | TBD |
| 1 | 112.038 | 4.816 | 30 000 | ~28 h | 163.5 s |
| 2 | 112.871 | 4.918 | 30 000 | ~28 h | 160.3 s |
| 3 | 113.163 | 4.850 | 30 000 | ~28 h | 160.9 s |

三个 seed 的 eval `length_mean = 1000.0`（play_env `episode_length_s =
20 s` × 50 Hz）。`success_rate` 为 `null`。
Expand All @@ -172,9 +172,9 @@ GPU vectorized 慢很多。

| Seed | 最终 `return_mean` | `return_std` | 收敛 iter | 训练 wall-clock | Eval wall-clock |
|---|---|---|---|---|---|
| 1 | TBD | TBD | 30 000 | TBD | TBD |
| 2 | TBD | TBD | 30 000 | TBD | TBD |
| 3 | TBD | TBD | 30 000 | TBD | TBD |
| 1 | 138.444 | 0.006 | 30 000 | ~28 h | 227.3 s |
| 2 | 137.980 | 0.008 | 30 000 | ~29 h | 301.6 s |
| 3 | 138.060 | 0.004 | 30 000 | ~29 h | 236.9 s |

Eval `length_mean = 1500.0`。Tracking play_env 默认 `episode_length_s =
1e9` 是为 viewer 无限 playback 设的；`genelab eval` 把它 clamp 到 30 s，
Expand All @@ -194,13 +194,15 @@ Eval `length_mean = 1500.0`。Tracking play_env 默认 `episode_length_s =

| Seed | 最终 `return_mean` | `return_std` | `success_rate` | 收敛 timestep | 训练 wall-clock | Eval wall-clock |
|---|---|---|---|---|---|---|
| 1 | TBD | TBD | TBD | 2 000 000（预算上限）| TBD | TBD |
| 2 | TBD | TBD | TBD | 2 000 000（预算上限）| TBD | TBD |
| 3 | TBD | TBD | TBD | 2 000 000（预算上限）| TBD | TBD |
| 1 | −6.122 | 13.307 | 0.990 | 2 000 000（预算上限）| ~67 min | 10.7 s |
| 2 | −7.260 | 14.695 | 0.990 | 2 000 000（预算上限）| ~62 min | 10.4 s |
| 3 | −8.790 | 18.275 | 0.970 | 2 000 000（预算上限）| ~64 min | 11.1 s |

Eval `length_mean = 100.0`（固定 episode length）。`success_rate` 由
manipulation 任务的 goal-reach termination 给出；每个 seed 的均值和三 seed
之间的总体均值在 v1.0 跑完之后回填。
manipulation 任务的 goal-reach termination 给出。三 seed 平均
`success_rate ≈ 0.983 ± 0.012` —— 比 v0.4.7 那次明显更紧凑：v0.4.7 有一
个 seed 因为末端朝向 drift 只到 0.89；v1.0 三个 seed 全部落在同一个
紧凑区间内。

!!! note "Previous: v0.4.7"

Expand Down