|
1 | 1 | Changelog |
2 | 2 | --------- |
3 | 3 |
|
| 4 | +0.5.1 (2026-05-22) |
| 5 | +~~~~~~~~~~~~~~~~~~ |
| 6 | + |
| 7 | +Added |
| 8 | +^^^^^ |
| 9 | + |
| 10 | +* Added an ``env_cfg`` block to ``teleop_replay_agent.py``'s stats output |
| 11 | + capturing the performance- and frame-timing-relevant env config inputs |
| 12 | + (``sim.dt``, ``sim.render_interval``, ``decimation``, ``episode_length_s``, |
| 13 | + ``scene.num_envs``, ``sim.device``, ``sim.use_fabric``, |
| 14 | + ``sim.render.antialiasing_mode``) along with precomputed ``policy_dt_s``, |
| 15 | + ``render_dt_s``, ``renders_per_step``, ``target_policy_hz``, and |
| 16 | + ``target_render_hz`` rates. The same fields are echoed in a compact |
| 17 | + ``Env timing:`` line in the stdout summary so the measured |
| 18 | + ``cpu_frame_time_ms`` / ``fps`` numbers are self-interpreting across |
| 19 | + machines and configs without cross-referencing the env definition. |
| 20 | + |
| 21 | +Changed |
| 22 | +^^^^^^^ |
| 23 | + |
| 24 | +* Changed ``teleop_replay_agent.py``'s ``cpu_frame_time_ms`` and ``fps`` |
| 25 | + blocks (both per-run and aggregate) to report on a **per-render** basis |
| 26 | + rather than per-``env.step``: each captured ``env.step`` CPU sample is |
| 27 | + divided by ``decimation / render_interval`` (the number of Kit renders |
| 28 | + per ``env.step``) before stats are computed. ``cpu_frame_time_ms.mean`` |
| 29 | + now reads as the wall time between rendered frames and ``fps.mean`` |
| 30 | + reads as the render rate -- the same number Kit's HUD shows, which is |
| 31 | + what the headset wearer / spectator actually perceives during real-time |
| 32 | + teleop. Field shapes and ``schema_version`` are unchanged. Falls back |
| 33 | + to the raw per-``env.step`` units when ``decimation`` or |
| 34 | + ``render_interval`` are unavailable from the env config. |
| 35 | + |
| 36 | + |
4 | 37 | 0.5.0 (2026-05-20) |
5 | 38 | ~~~~~~~~~~~~~~~~~~ |
6 | 39 |
|
|
0 commit comments