sensors-visualization
Configure SMARTS sensor observations, optional Panda3D rendering, Envision recording and replay, and safe diagnostics without conflating CPU checks with graphics services.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- sensors-visualization
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Configure SMARTS sensor observations, optional Panda3D rendering, Envision recording and replay, and safe diagnostics without conflating CPU checks with graphics services.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/data-formats.md4.3 KB
- 参考文档references/envision-and-replay.md4.4 KB
- 参考文档references/rendering-and-cameras.md4.8 KB
- 参考文档references/sensor-reference.md5.3 KB
- 参考文档references/troubleshooting.md5.3 KB
- 脚本scripts/check_rendering.py3.6 KB
- 脚本scripts/inspect_replay_records.py3.1 KB
03 SKILL.md 正文
SMARTS sensors and visualization
Use this route when an agent needs additional SMARTS observations, camera/grid
outputs, custom shader renders, Envision state streaming or replay, or a
read-only check of rendering and replay prerequisites. Keep the simulation
lifecycle and environment creation in simulation-environments; use
cli-integrations for exact scl command syntax. Frontend npm development,
raw media, and long-running servers are outside this route.
Decide which boundary you are testing
- Start with a CPU/non-rendered interface when the policy only needs state, waypoints, lidar, lane positions, accelerometer, signals, or nearby actors.
- Treat
drivable_area_grid_map,occupancy_grid_map,top_down_rgb, andcustom_rendersas software-rendering features. They require the camera extra and a usable Panda3D/OpenGL offscreen or X11 display. - An
occlusion_mapalso requires an occupancy map with identical width and height. It is a rendered output, not a CPU-only visibility calculation. - Envision is a separate client/server/browser path. A successful SMARTS import or Panda3D import does not prove that the server is listening.
Read the linked references before building an interface:
- sensor-reference.md — configuration dataclasses, sensor semantics, and observation contracts.
- rendering-and-cameras.md — camera extras, offscreen rendering, custom shaders, and performance boundaries.
- envision-and-replay.md — client/server, JSONL recording, replay, and Visdom limits.
- data-formats.md — shapes, metadata, and stable serialization expectations.
- troubleshooting.md — failures by install, configuration, API misuse, renderer, and service workflow.
Configure the interface
Use AgentInterface with explicit dataclass values when a non-default shape,
range, or lookahead matters. Boolean True resolves to the corresponding
config dataclass; False disables the sensor. The interface defaults
accelerometer and lane positions on, but leaves most expensive or optional
sensors off. AgentInterface.from_type(AgentType.Full, ...) is convenient but
also enables rendered observations and lidar; do not use it as a cheap smoke
interface.
from smarts.core.agent_interface import (
AgentInterface, Waypoints, NeighborhoodVehicles, RGB, OGM,
)
interface = AgentInterface(
waypoint_paths=Waypoints(lookahead=32),
neighborhood_vehicle_states=NeighborhoodVehicles(radius=50),
top_down_rgb=RGB(width=128, height=128, resolution=100 / 128),
occupancy_grid_map=OGM(width=128, height=128, resolution=100 / 128),
action=..., # select an action type in the environment route
)
Before stepping, inspect the environment's per-agent observation space and
confirm that the policy reads only enabled fields. After a step, check None
for disabled optional fields and preserve the named observation structures;
do not assume every observation is a flat NumPy array.
Workflow and checks
- For CPU checks, import
AgentInterface, construct low-dimensional sensor configs, and run a bounded environment test through the sibling lifecycle route. This does not require Panda3D. - For camera checks, run
scripts/check_rendering.py, preferably underxvfb-run -a. Use--probe-offscreenonly for a bounded construction and teardown probe; it is not a renderer behavior test. - For Envision records, use
scripts/inspect_replay_records.pyon a file or directory. It never sends data, starts a server, or modifies files. - Keep image dimensions and resolution small while diagnosing. Image sensors
can dominate
step()time and memory. - For recording or replay, verify the server endpoint separately and inspect the generated JSONL before attempting playback. Use the CLI sibling route for command spelling; this skill documents the Python client contract only.
Verified limits
The prepared SMARTS 2.0.1 environment passed package imports, live interface and lidar signature inspection, and a Panda3D/SMARTS renderer import smoke under Xvfb. That is evidence for importability only. Full renderer behavior, large image observations, Visdom, and a live Envision server remain optional or unverified. The bundled helpers report these limits rather than claiming a successful renderer or visualization session.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).