video2sim

← 教程资料库 · 仿真与基准

本节介绍 LeRobot 的通用功能。请按目标机器人、数据集和运行环境配置示例；用于 AlohaMini 2 / 2 Pro 前，需要确认硬件与接口兼容性。

---

Phone video → photoreal Isaac Sim scene with a physics collider, on an 8 GB GPU.

video2sim turns a single hand-held phone walkthrough into an Isaac Sim 5.x stage: a NuRec Gaussian-splat room you can render photorealistically, plus an invisible, metric, gravity-aligned TSDF collider the robot can actually touch.

It is SfM-free: no COLMAP/GLOMAP mapping in the loop. Camera poses come from LingBot-Map streaming inference, get jointly refined during splat training (per-view se3 through the rasterizer gradients, view 0 anchored), and the splat is initialized densely from the MV-consistent TSDF cloud — InstantSplat-style. That dense init is what makes the whole thing fast: 10k iterations ≈ 20 min end-to-end training instead of the 80k-iteration growth schedule it replaces.

Every default in this package is a hard-won fix from ~10 failed runs on the validated room3 capture. The defaults of record live in configs/default.yaml — do not "improve" them without re-validating.

---

Architecture

Each stage is one module: importable and runnable as python -m video2sim.<stage>. The orchestrator (python -m video2sim run) sequences them as subprocesses, routing each stage to the interpreter its contract demands (it never auto-switches inside a process).

 phone video (.mp4)
      |
      v
 [extract]        video -> frames/f_%04d.jpg  (ffmpeg, fps 8)      | main venv (Basic_RL)
      |
      v
 [lingbot_infer]  frames -> lingbot/pred-*.pt                      | main venv
      |           streaming poses + depth; fp8 KV + sliding        |  + PYTHONPATH=<lingbot fork>
      |           window 48 (the 8 GB recipe)                      |  + ninja on PATH
      +---------------------------+
      |                           |
      v                           v
 [fuse]                       [to_colmap]                          | main venv
  MV-consistency filter +      LingBot E/K + TSDF cloud            |
  4 mm TSDF fusion             -> COLMAP text sparse/0             |
  -> fuse/mvtsdf.ply(+mesh)    (dense 350k init, NO SfM)           |
      |                           |
      v                           v
 [refine]                     [train]                              | main venv, GPU EXCLUSIVE
  island filter, RANSAC        gsplat MCMC 10k iters,              |
  floor -> z=0, metric         packed=True, D-SSIM, PPISP-lite,    |
  scale (ceiling 2.4 m)        joint pose-opt (view 0 anchor)      |
  -> final_mesh/cloud/align    -> splat.pt (+refined_poses.npz)    |
      |                           |
      v                           v
 [scene_prep]                 [export]                             | main venv;
  decimate to 300k tris        splat.pt -> SH3 zero-padded PLY     |  USDZ substep runs under
  -> collider.npz              -> free-space prune (TSDF oracle)   |  nurec-venv (subprocess)
      |                        -> export/scene.usdz                |
      +------------+--------------+
                   v
              [scene]   splat (align transform) + invisible        | isaac5-venv
                        collider + robot + dome 450
                        -> scene/scene.usd            [--gui]

 side tools:  [spots]  open-spot finder + capture-camera pick      | main venv
              [check_env]  preflight pass/fail table               | any python (probes venvs)

Interpreter contract (documented in every module docstring, never auto-switched):

Stage(s)	Interpreter
extract, lingbot_infer, fuse, refine, to_colmap, train, export (steps 1–2), scene_prep, spots	/home/perelman/Basic_RL/.venv/bin/python
export step 3 (PLY → NuRec USDZ, _nurec_convert.py)	/home/perelman/nurec-venv/bin/python
scene (Isaac USD assembly)	/home/perelman/isaac5-venv/bin/python
LingBot child process	main venv + PYTHONPATH=<lingbot fork> + ninja on PATH (the wrapper sets both)

---

Requirements

• GPU: NVIDIA, 8 GiB VRAM minimum (validated card: RTX 4060 8 GB — the whole recipe is 8 GB-bounded). The GPU must be exclusive during LingBot inference, CUDA SfM and splat training: an open Isaac GUI alone eats 3–4 GB and OOMs the run.

• RAM: 32 GB (validated on a 31 GB box; the trainer caches images as uint8 ≈ 3 GB — float32 would be ~12 GB — and fuse warns at 45 GB RSS).

• Binaries: ffmpeg / ffprobe on PATH.

• Venv inventory:

Venv	Path	Must provide
main	/home/perelman/Basic_RL/.venv	torch + CUDA, gsplat ≥ 1.5 (MCMCStrategy + packed rasterization), pycolmap, open3d, scipy, PIL, ninja in bin/ (the fp8-KV CUDA extension is JIT-compiled)
NuRec	/home/perelman/nurec-venv	pxr (USD), torch, msgpack, ncore — for the threedgrut export path
Isaac	/home/perelman/isaac5-venv	isaacsim 5.x + pxr (incl. omni.usd.schema.omni_nurec_types)

• Checkouts / models: LingBot-Map RTX4060-8GB fork (with scripts/predict_stream.py), LingBot checkpoint lingbot-map-long.pt, and a 3dgrut checkout (/home/perelman/3dgrut) for the NuRec exporter modules.

Setup

Run the preflight check first — it verifies every runtime the pipeline touches before you burn an hour on a run (venv imports in subprocesses, ffmpeg, LingBot fork + ckpt, ninja, GPU VRAM/exclusivity, RAM), and exits nonzero on any FAIL:

cd /home/perelman/AlohaMini/video2sim
python3 -m video2sim.check_env

check_env is stdlib-only and runs under any Python; heavy imports are probed in subprocesses under their own interpreters. Read the WARN rows too — a "busy GPU" warning means training/inference will OOM if you proceed.

Quickstart

One command, end-to-end (idempotent — re-running skips stages whose outputs exist):

cd /home/perelman/AlohaMini/video2sim
/home/perelman/Basic_RL/.venv/bin/python -m video2sim run ~/room.mp4 --workdir runs/room1

Then open the assembled scene in the Isaac GUI (needs the GPU free again):

/home/perelman/Basic_RL/.venv/bin/python -m video2sim run --workdir runs/room1 --gui

Outputs land in the workdir: frames/, lingbot/pred-*.pt, fuse/mvtsdf.ply(+_mesh), final_{mesh,cloud}.ply + final_align.npz, sparse/0/, splat.pt, export/scene.usdz, collider.npz, scene/scene.usd.

Per-stage overrides go through a YAML config (--config, key stage_args) or by running any single stage directly, e.g. python -m video2sim.train --workdir runs/room1 --iters 20000 --video. For a ~60 s capture the wall-clock is ≈ 30–40 min, training-dominated (~20 min).

The standard recipe (and WHY each number is what it is)

Mirrors configs/default.yaml == the argparse defaults. Every WHY below was measured on this hardware; the odd-looking constants are failure post-mortems, not taste.

Stage	Setting	WHY (measured)
extract	fps=8, -q:v 2	Dense enough for sequential matching at walking speed, sparse enough to stay tractable. Thinning happens HERE — downstream is stride 1 always. Near-lossless JPEG: feature extractors are quality-sensitive.
extract	autorotate, no transpose filter	ffmpeg applies rotation metadata before -vf; adding transpose double-rotates. Orientation judged on DISPLAY size.
(any SfM validation)	--single-camera MANDATORY	Per-frame independent cameras collapse focal lengths to −1e6…+8e4 → training loss oscillation + fog. Shared camera recovers fx 1642 (independently matching LingBot's 1643).
lingbot_infer	fp8 KV cache + sliding-window 48	THE 8 GB recipe. sw64+fp8 = confirmed OOM (KV +0.6 GB collides with the attention peak).
lingbot_infer	max_frame_num = actual frames + 8	FlashInfer pre-allocates the paged-KV pool from it; the fork's default 1024 was 6.7 GB on its own → OOM on 8 GB.
lingbot_infer	image_size stays 518	The ckpt pos_embed (1,1370,1024) is size-bound — never override.
fuse	voxel 4 mm, trunc 2 cm, conf > 2.3, neighbors ±4/8/12, ≥ 2 consistent	Collider-grade surface that still fits RAM. Kept-pixel % doubles as a pose-health metric: same-track ~88 %, cross-track collapses to ~58 %.
refine	camera-track gates on floor RANSAC (above > 0.9, med > 0.2, spread < 0.45)	The biggest plane is often a WALL. The spread gate is what rejects walls: a hand-held phone keeps near-constant height over the FLOOR but varying distance to walls.
refine	ceiling anchor 2.4 m, fallback hand-height 1.35 m	Metric scale from floor→ceiling when a horizontal ceiling plane is visible; else phone hand height over the floor.
to_colmap	dense init 350k points from the TSDF cloud	The InstantSplat fast path: growth-free training converged in 10k iters vs 80k, with 38× fewer floaters. LingBot extrinsics are world→cam in COLMAP convention — qvec/tvec written directly, no inversion.
train	iters=10000, cap=350000, downscale=1	10k suffices because of the dense init. cap 500k = measured OOM; 350k @1080p fits. (Yesterday's 1080p OOM streak was actually the extra depth render channel, not the resolution.)
train	packed=True ALWAYS	Sparse tile-intersection buffers are the difference between OOM and full-res on 8 GB (unpacked died at ~260k gaussians even at 2× downscale).
train	lrs keyed BY NAME	torch.nn.ParameterDict sorts keys alphabetically; a positional zip silently shuffled every lr (opacities got 1e-3 instead of 5e-2) — the single biggest cause of the fog splats.
train	scene_scale capped at 6.0	Corridor/forward-motion scenes give a huge camera spread that over-inflates means-lr + MCMC noise and collapses opacity to fog.
train	D-SSIM 0.2 term	Plain L1 is minimized just as well by big translucent blobs — the fog optimum.
train	joint pose-opt on, lr 1e-4, view 0 anchored	Absorbs feedforward pose drift without SfM: LingBot poses measured at 54 cm RMS vs a GLOMAP/BA trajectory (Umeyama residual; reproduced 53.8). Corrections keep growing with budget (1 cm median @20k → 3.8 cm @80k) — joint opt asymptotes toward SfM.
train	PPISP-lite per-view exposure/WB	Phone auto-exposure/AWB otherwise gets baked into the splat as geometric mud; identity at novel-view time.
train	--depth-reg SAME-track poses only	LingBot depth + poses from a different track = fog hedging (every gaussian goes translucent); MV re-fusion drops 88 % → 58 %. Depth and poses are one body.
train	optimizer step first, THEN MCMC relocation	Official ordering; relocating first applied stale grads to teleported gaussians.
export	SH3 zero-padded PLY (62 props)	NuRec assumes radiance_sph_degree=3; an SH0-only payload hits a pathological ~47 s/frame slow path. Zero f_rest is visually identical to SH0.
export	free-space prune, radius 0.20 m, spatial criterion ONLY	TSDF cloud is the surface oracle; conservative on purpose (TSDF holes only lose gaussians beyond the radius). Opacity pruning is FORBIDDEN for MCMC ensembles — min_opacity cuts deleted ~79 % of the splat (MCMC keeps meaningful low-opacity gaussians).
scene_prep	collider budget 300k tris	Full contact honesty without blowing up PhysX cooking (same budget as the CuRobo env-mesh recipe).
scene	dome intensity 450, no key light	The mesh-room recipe (dome 1200 + key 2000) overexposes the NuRec splat to white blur; approved GUI views used ~350–450.
scene	collider identity transform; only the splat gets the align matrix	The refined mesh is already metric/aligned (floor z=0); the splat still lives in the raw LingBot frame. Gf.Matrix4d is row-vector (p' = p @ M).
spots	candidates within 1.2 m of the walked track	The global max-clearance cell was reconstruction garbage OUTSIDE the walls — noise past the room boundary looks like open floor to a naive distance transform.
spots	demo camera = a real capture-trajectory pose	A synthetic camera 2.4 m out sat inside splat furniture and filmed only blur; a pose the videographer actually stood at is guaranteed fog-free.

Troubleshooting

Failure modes actually hit during development, symptom-first:

• Splat renders as fog / translucent blobs. Check in order: (1) lrs keyed by name, not position (ParameterDict alphabetical-sort trap); (2) D-SSIM term present; (3) scene_scale capped (corridor scenes); (4) you did NOT pair --depth-reg with poses from a different track.

• Training OOM on 8 GB. packed=True must be on; cap ≤ 350k (500k OOMs); drop --depth-reg (the extra RGB+ED render channel is what killed 1080p runs); make sure nothing else holds VRAM — the Isaac GUI alone eats 3–4 GB.

• LingBot OOM in the FlashInfer paged-KV pool. max_frame_num must be sized to the sequence (the wrapper does this; the fork default 1024 pre-allocates 6.7 GB). Sliding window > 48 with fp8 also OOMs. GPU must be exclusive.

• ninja: not found during LingBot startup. The fp8-KV CUDA extension is JIT-compiled and needs ninja in the main venv bin (leads PATH in the child); check_env verifies this.

• NuRec renders at ~47 s/frame. Your PLY is SH0-only. Export through video2sim.export — it zero-pads the 45 f_rest SH3 props.

• Splat mostly disappeared after export. Opacity pruning was applied to an MCMC ensemble. Keep --min-opacity 0.0 (the default); pruning is spatial-only.

• Scene washes out to white blur in Isaac. Dome intensity too high for a NuRec splat — use the default 450, no key light.

• fuse kept-pixel % collapses (~15–17 %). Capture problem, not code: portrait orientation + fast pans (room2) vs landscape + slow walk (room3, 88 %). Re-shoot landscape and slow; kept-% is the health gate to check first.

• Floor/scale wrong after refine. A wall won the RANSAC or no ceiling was seen: check the per-plane gate log (above/med/spread), and whether the hand-height fallback (1.35 m) fired on a scene where the ceiling anchor should have.

• extract refuses to run. Stale f_*.jpg in the frames dir — they silently poison the single-camera track (mixed resolutions / numbering). Pass --force to clobber deliberately.

• Open-spot picker chose a point outside the room. Keep the walked-track restriction (--track-radius 1.2); the unrestricted max-clearance point was recon garbage beyond the walls.

• Demo camera films only blur. It was placed at a synthetic viewpoint inside splat geometry — pick a real capture pose (video2sim.spots --camera-target ...).

• Stage crashes mid-pipeline. The runner is resumable: re-run the same python -m video2sim run command; stages with existing outputs are skipped (--force to redo, --from/--until to slice).

Results — v10 vs v11

Both validated on the room3 capture (62 s landscape 1080p phone walkthrough):

	v10	v11 (standard)
Init	sparse points, MCMC growth	dense 350k TSDF init (InstantSplat-style)
Iters	80k (most of the gain harvested in the post-relocation settle phase)	10k
Training wall-clock	~7× v11 (hours)	~20 min
Floaters	baseline	38× fewer (vs sparse init)
Visual	slightly better (user judgment)	approved; not worth 7×

v11 is the adopted standard — this package's defaults are v11. Use the v10 budget (--iters 80000 + growth) only when squeezing the last bit of sharpness matters more than turnaround. Reference artifacts from the session: room3_nurec_v10.usdz / room3_nurec_v11.usdz (+ matching room3_splat_v*.ply).

Repository layout

video2sim/
  __main__.py       python -m video2sim -> cli
  cli.py            end-to-end runner (subprocess per stage, resume, venv routing)
  check_env.py      stage 0: preflight pass/fail table
  extract.py        stage 1: video -> frames (ffmpeg)
  lingbot_infer.py  stage 2: frames -> pred-*.pt (LingBot 8GB recipe)
  fuse.py           stage 3: MV-consistency filtered TSDF fusion
  refine.py         stage 4: island filter + gravity/metric alignment
  to_colmap.py      stage 5: LingBot poses + TSDF cloud -> COLMAP text model
  train.py          stage 6: gsplat MCMC trainer (+ joint pose-opt, flythrough)
  export.py         stage 7: splat.pt -> SH3 PLY -> free-space prune -> NuRec USDZ
  _nurec_convert.py    (NuRec-venv helper subprocess for the USDZ step)
  scene_prep.py     stage 8: refined mesh -> decimated collider npz
  scene.py          stage 9: splat + collider + robot -> scene.usd (Isaac venv)
  spots.py          open-spot finder + capture-camera pick
configs/default.yaml  the recipe of record (documented defaults)
tests/test_smoke.py   CPU-only smoke tests

Tests

CPU-only smoke tests (se3 exp vs scipy, 62-prop PLY roundtrip, COLMAP text ↔ pycolmap, quaternion roundtrips, free-space pruning):

cd /home/perelman/AlohaMini/video2sim
/home/perelman/Basic_RL/.venv/bin/python -m pytest tests/ -q