Problem
RenderContext publishes renderer scene state at most once for a physics-step
count. An environment reset can change asset state and call forward() without
advancing that count, but the environment reset paths do not invalidate the
scene-state cadence. The first camera read after reset can therefore skip the
official renderer transform update and reuse pre-reset transforms.
This affects all reset implementations that own _reset_idx():
ManagerBasedEnvManagerBasedRLEnv (which overrides the manager base implementation without
calling super())DirectRLEnvDirectMARLEnv
Minimal reproduction
The lifecycle defect is renderer-independent:
env.sim.render_context._last_scene_state_step = 7
env.reset()
assert env.sim.render_context._last_scene_state_step is None
On current develop, the value remains 7 for all four environment types.
Temporarily removing the proposed fix makes this regression fail with
assert 7 is None.
Newton/OVRTX evidence
The issue was originally reproduced with Isaac Lab
v3.0.0-beta2.patch1, isaaclab_newton==0.13.6, Newton 1.2.1,
isaaclab_ov==0.4.2, and ovrtx==0.3.0.312915:
- Before reset, every audited official OVRTX body matrix matched Newton
body_q with maximum absolute error 1.67e-6.
- After a same-step reset without cadence invalidation, Newton body state and
camera pose were correct, but OVRTX matrix error grew to about 1.79--1.99
and a non-temporal depth AOV retained the displaced pose.
- Calling only the public patch1
RenderContext.reset_transform_cadence() restored matrix error to 1.67e-6.
- With cadence invalidation, baseline versus first post-reset depth MAE was
2.33e-7; displaced versus post-reset depth MAE was about 0.484.
No custom transform publication, stage rebuild, native OVRTX history reset,
CPU-transform fallback, or warm-up was used. Native Renderer.reset() is not
part of the solution: it restarts RTPT accumulation and makes immediate beauty
frames intentionally cold, while the non-temporal depth test proves cadence
invalidation alone fixes the stale geometry.
Expected behavior
After environment state has been reset, the next renderer update must publish
scene state even when the physics-step count did not advance.
Proposed fix
Call RenderContext.reset_scene_state_cadence() at the end of each environment
reset implementation listed above and add backend-neutral regression coverage
for all four paths. This keeps transform ownership in RenderContext and the
official renderer adapters and does not introduce renderer-specific lifecycle
hooks.
Problem
RenderContextpublishes renderer scene state at most once for a physics-stepcount. An environment reset can change asset state and call
forward()withoutadvancing that count, but the environment reset paths do not invalidate the
scene-state cadence. The first camera read after reset can therefore skip the
official renderer transform update and reuse pre-reset transforms.
This affects all reset implementations that own
_reset_idx():ManagerBasedEnvManagerBasedRLEnv(which overrides the manager base implementation withoutcalling
super())DirectRLEnvDirectMARLEnvMinimal reproduction
The lifecycle defect is renderer-independent:
On current
develop, the value remains7for all four environment types.Temporarily removing the proposed fix makes this regression fail with
assert 7 is None.Newton/OVRTX evidence
The issue was originally reproduced with Isaac Lab
v3.0.0-beta2.patch1,isaaclab_newton==0.13.6, Newton1.2.1,isaaclab_ov==0.4.2, andovrtx==0.3.0.312915:body_qwith maximum absolute error1.67e-6.camera pose were correct, but OVRTX matrix error grew to about
1.79--1.99and a non-temporal depth AOV retained the displaced pose.
RenderContext.reset_transform_cadence()restored matrix error to1.67e-6.2.33e-7; displaced versus post-reset depth MAE was about0.484.No custom transform publication, stage rebuild, native OVRTX history reset,
CPU-transform fallback, or warm-up was used. Native
Renderer.reset()is notpart of the solution: it restarts RTPT accumulation and makes immediate beauty
frames intentionally cold, while the non-temporal depth test proves cadence
invalidation alone fixes the stale geometry.
Expected behavior
After environment state has been reset, the next renderer update must publish
scene state even when the physics-step count did not advance.
Proposed fix
Call
RenderContext.reset_scene_state_cadence()at the end of each environmentreset implementation listed above and add backend-neutral regression coverage
for all four paths. This keeps transform ownership in
RenderContextand theofficial renderer adapters and does not introduce renderer-specific lifecycle
hooks.