TAO CenterPose Pose Estimation Skill
Guides agents through training, evaluation, export, and deployment of CenterPose models for 6-DoF object pose estimation.
The skill restricts its declared tools to Read and Bash and documents Docker/NVIDIA Container Toolkit requirements, data paths, checkpoint mappings, and deployment templates. However, training, evaluation, export, and TensorRT operations create external compute and file effects, while stepwise confirmation, path isolation, sensitive-data/key handling, rollback, and recovery are not specified, so points are deducted.
Actions, inputs and outputs, dataset paths, checkpoint inference, and several error patterns are described coherently, with deterministic settings noted. Still, there are no executable tests covering key paths; schema/template availability depends on packaging state; AutoML routing has conditional branches; and abnormal-input or dependency-failure diagnostics are thin, so points are deducted.
Trigger phrases, CenterPose training/evaluation/export/inference scenarios, single-node limits, GPU requirements, fixed keypoint count, and some non-fit conditions are reasonably clear. Chinese-language interaction, TAO/container version boundaries, and mainland-China network reachability are not addressed, and NVIDIA container-image access may constrain the environment, so points are deducted.
Front matter, Apache-2.0 licensing, version 0.1.0, author, tags, reference templates, parameter tables, error patterns, and a benchmark report provide a useful maintenance base. However, the recommended Instructions/Examples/FAQ structure is absent, there is no changelog or explicit skill maintenance/update path, and the benchmark flags the author format as nonconforming, so points are deducted.
The documentation covers dataset preparation, training, evaluation, inference, export, TensorRT, and checkpoint handoff, with templates and parameters that can directly guide the core workflow. But no representative usable artifacts or independent result evidence are statically verifiable; correctness, completeness, and marginal benefit over manual TAO use require execution, so the static cap applies and points are deducted.
The skill includes auditable YAML templates, action mappings, error patterns, and a limited benchmark report, supporting traceability for some claims. However, the benchmark has only one task, there is no skill-specific committed test suite or real CI coverage, and the reported results cannot be independently reproduced from the supplied files alone, so points are deducted.
- Before execution, confirm the authorized input/output directories, checkpoints, and encryption-key handling; constrain Bash and container access to the task scope.
- Verify the actual TAO, Docker, NVIDIA Container Toolkit, image, and schema/template versions and availability; the files do not establish direct dependency reachability from mainland-China networks.
- A successful TensorRT engine generation does not establish deploy evaluate/inference success; inspect exit codes and logs separately and retain recoverable outputs and checkpoints.
What does this skill do, and when should you use it?
This skill supports NVIDIA TAO Toolkit workflows for CenterPose keypoint and pose estimation. It covers training, evaluation, inference, ONNX export, TensorRT engine generation, and deployment. CenterPose detects object centers and regresses keypoint locations for 6-DoF object pose estimation. It requires Docker, NVIDIA Container Toolkit, and GPU hardware, and defines rules for data paths, AutoML, checkpoints, inference mappings, and deployment templates.
Reads packaged TAO schemas, specification templates, and skill_info configuration; constructs action-specific paths for training, validation, test, inference, and TensorRT calibration data; guides CenterPose training, evaluation, inference, ONNX export, and TensorRT engine generation; applies AutoML policy and default routing; resolves upstream checkpoints, ONNX files, calibration caches, TensorRT engines, and result directories; applies mappings for encryption keys, pretrained backbones, and resume checkpoints; and addresses archive extraction, keypoint-count mismatches, and TensorRT deployment postprocessor errors.
- A computer-vision engineer wants to train a CenterPose model on custom centerpose-format data.
- A model developer needs to evaluate a training checkpoint using separate validation or test data.
- A deployment engineer wants to export CenterPose to ONNX and generate an FP32, FP16, or INT8 TensorRT engine.
- A robotics or industrial-vision team needs 6-DoF pose inference using PnP and camera intrinsics.
- A TAO user wants AutoML enabled by default or disabled for one training run.
What are this skill's strengths and limitations?
- Covers the main TAO workflow from CenterPose training through TensorRT deployment.
- Documents per-action data sources, important parameters, checkpoint inference, and error patterns.
- Supports AutoML run policy and documents fixed input resolution, keypoint count, and export settings.
- Packages deployment templates and generated schemas for supported actions.
- Requires Docker, NVIDIA Container Toolkit, and at least one GPU with 16GB or more VRAM.
- Supports single-node configuration only; no explicit num_nodes or distributed_strategy settings are provided.
- Users must prepare and extract train.tar.gz, val.tar.gz, and test.tar.gz archives.
- The source provides no specific platform test matrix, performance benchmarks, or cost information.
- The skill declares Apache-2.0, while the repository README describes the collection as dual-licensed Apache-2.0 and CC BY-4.0; the applicable scope should be checked separately.
How do you install this skill?
Install it with the skills CLI command documented in the README: npx skills add nvidia/skills --skill tao-train-centerpose --yes. The CLI installs the selected skill from the NVIDIA/skills collection; the source does not specify the exact destination directory.
How do you use this skill?
After installation, submit a focused request in an Agent Skills-compatible client, such as “train CenterPose” or “run 6-DoF object pose inference.” Provide Docker, NVIDIA Container Toolkit, GPU access, and the extracted data directories required by the selected action. Training defaults to AutoML; phrases such as “disable AutoML” or “plain training” apply an off policy for that run.