TAO PointPillars 3D Detection
Guides TAO workflows for training, evaluating, and deploying LiDAR-based PointPillars detectors.
The skill limits declared tools to Read and Bash and requires Docker plus NVIDIA Container Toolkit. It discloses dataset paths, checkpoints, result directories, encryption keys, container mounts, and warns against reusing paths from another results root. Deductions apply because it does not define user confirmation, sensitive-data handling, least-privilege Bash boundaries, rollback, or recovery procedures; S3 paths imply external data flow without complete controls. No malware, credential theft, covert exfiltration, or destructive default is shown.
SKILL.md provides action mappings, dataset prerequisites, path rules, multi-node variables, common failure patterns, and status/artifact checks, making the happy path reasonably clear from static evidence. Deductions apply because there is no executable reproduction evidence or committed key-path test coverage; skill_info.yaml conflicts with the document on the inference data source, the referenced deployment document is not included in the supplied material, and abnormal cases still rely heavily on agent judgment. Static calibration caps this at 10.
Trigger phrases, audience scenarios, training/evaluation/export/pruning/inference scope, GPU requirements, and data format are fairly explicit, and the document states that parent-model gen_trt_engine and a separate retrain subcommand are unsupported. Deductions apply because non-fit boundaries and input/output contracts remain incomplete, Chinese-language interaction is not addressed, and reachability of Docker images, S3, or NVIDIA-related services from mainland China is not established. The local GPU/Docker dependency also limits environment fit.
The material includes frontmatter, action-oriented sections, schema/template references, parameter examples, error patterns, a version, and an author. BENCHMARK.md records an evaluation date and maintenance recommendation, while the repository README describes daily mirroring. Deductions apply because recommended Instructions and Examples sections are absent, the benchmark reports folder-depth, schema-location, and author-format findings, and the skill does not clearly state changelog, maintenance ownership, or an update path. The supplied license metadata is NOASSERTION while the files state Apache-2.0, requiring governance clarification.
The skill covers PointPillars dataset conversion, training, resumption, evaluation, export, pruning, retraining, and inference, with directly reusable spec_overrides examples, artifact names, and failure criteria. Deductions apply because static review cannot verify result correctness; the single-task benchmark reports 20% Codex correctness and 44% effectiveness, with no negative-trigger cases, so direct usability and generalization evidence are limited.
A pinned revision, multiple configuration templates, skill_info.yaml, one evaluation case, and BENCHMARK.md provide limited cross-file auditability. Deductions apply because evaluation contains only one task with one attempt, there is no committed key-path test suite or real CI execution evidence, and the benchmark PASS does not establish reproducibility for this static review. Static calibration therefore keeps this at 5 or below.
- Do not treat NVIDIA provenance or the benchmark PASS as guarantees of reliability, safety, or effectiveness; actual training and deployment reproducibility is unverified.
- Before execution, reconcile the data-source discrepancy between skill_info.yaml and SKILL.md, and verify that the current container can see dataset.data_info_path, dbinfos_train.pkl, infos_train.pkl, and infos_val.pkl.
- Training, pruning, export, and deployment may consume GPUs, write result directories, and process sensitive point-cloud data; obtain user confirmation, isolate the results root, and preserve recoverable copies.
- Confirm reachability of the Docker image, data sources, and related services in the target environment, especially from mainland-China networks.
- Reconcile the Apache-2.0/CC-BY-4.0 file declarations with the supplied NOASSERTION license metadata before publication or redistribution.
What does this skill do, and when should you use it?
This skill supports TAO Toolkit workflows for 3D object detection from LiDAR point clouds. It covers dataset conversion, training, evaluation, inference, export, pruning, and retraining from a pruned model, with separate TensorRT deployment guidance. PointPillars converts point clouds into a pillar-based pseudo-image before applying 2D detection. It requires Docker, NVIDIA Container Toolkit, and at least one GPU.
Reads the skill's reference files and model, dataset, checkpoint, and output path configuration; runs the packaged PointPillars CLI actions dataset_convert, train, evaluate, inference, export, and prune; and retrains pruned models with train.pruned_model_path. Training is AutoML-enabled by default and, when the required schema and template are packaged, routes through tao-skill-bank:tao-run-automl. TensorRT gen_trt_engine, TensorRT evaluation, and TensorRT inference require reading tao-deploy-pointpillars.md first. The workflow also checks converted data-info files, status.json, and expected model or ONNX artifacts.
- An autonomous-driving engineer wants to train PointPillars on extracted LiDAR data with the required train/lidar, train/label, val/lidar, and val/label structure.
- A robotics researcher needs to evaluate or run inference with an existing PointPillars checkpoint.
- A deployment engineer wants to export a PointPillars model to ONNX or prepare TensorRT deployment artifacts.
- A model-optimization engineer wants to prune a trained model and retrain it from the resulting encrypted .tlt artifact.
- A multi-GPU or multi-node user needs torchrun-based training with native PyTorch DistributedDataParallel.
What are this skill's strengths and limitations?
- Covers the main PointPillars lifecycle: conversion, training, evaluation, inference, export, pruning, and retraining.
- Documents strict requirements for conversion artifacts, checkpoint names, container mounts, and status verification.
- Supports AutoML, torchrun, multi-GPU, and multi-node training.
- Provides an entry point for TensorRT deployment guidance and explains packaged TAO schemas and templates.
- Requires Docker, NVIDIA Container Toolkit, and an NVIDIA GPU; 16GB or more VRAM per GPU is recommended.
- dataset_convert must run first, and data-info files must be visible inside the current container, making path and mount configuration sensitive.
- The supplied material provides no independent test suite, concrete dataset, or end-to-end success example.
- The repository README states Apache 2.0 plus CC BY 4.0, while this skill is marked Apache-2.0; adopters should distinguish the skill's license from the collection repository's license scope.
How do you install this skill?
Install the individual skill from NVIDIA's collection with: npx skills add nvidia/skills --skill tao-train-pointpillars --yes. The README does not document the exact installation directory for each client; after installation, the skill is available when the agent loads skills and encounters a relevant task.
How do you use this skill?
After installation, ask for a task such as “train PointPillars,” “LiDAR 3D detection,” or “point-cloud object detection.” Prepare Docker, NVIDIA Container Toolkit, a GPU, and an extracted PointPillars dataset. Run dataset_convert first, then provide dataset.data_info_path from the current results root for later actions. Supply the relevant checkpoint, data paths, and output paths for training, evaluation, inference, export, pruning, or retraining, and verify status.json plus the expected artifacts.