[Bug Report] TerminationManager._term_dones not updated per step (stale values during rollout) #3720
Description
Describe the bug
Hi, thank you for your excellent work.
I noticed that in the current implementation of TerminationManager.compute(), the internal buffer self._term_dones is only updated when a termination actually occurs, but not during the normal running steps of the environment.
Current Behavior
def compute(self) -> torch.Tensor:
"""Computes the termination signal as union of individual terms.
This function calls each termination term managed by the class and performs a logical OR operation
to compute the net termination signal.
Returns:
The combined termination signal of shape (num_envs,).
"""
# reset computation
self._truncated_buf[:] = False
self._terminated_buf[:] = False
# iterate over all the termination terms
for i, term_cfg in enumerate(self._term_cfgs):
value = term_cfg.func(self._env, **term_cfg.params)
# store timeout signal separately
if term_cfg.time_out:
self._truncated_buf |= value
else:
self._terminated_buf |= value
# add to episode dones
rows = value.nonzero(as_tuple=True)[0] # indexing is cheaper than boolean advance indexing
if rows.numel() > 0:
self._term_dones[rows] = False
self._term_dones[rows, i] = True
# return combined termination signal
return self._truncated_buf | self._terminated_bufHere, self._term_dones is only modified when rows.numel() > 0, i.e., when a specific termination condition is triggered.
During the rest of the rollout (normal running steps), self._term_dones keeps the values from the previous episode, and is therefore stale.
Consequences
For logging (e.g. inside reset()), this behavior is harmless — it still correctly summarizes the last episode via:
last_episode_done_stats = self._term_dones.float().mean(dim=0)However, for runtime queries like:
get_term(name)
get_active_iterable_terms(env_idx)the returned tensors will not reflect the current state of termination terms, but rather the previous episode’s data.
This can be problematic if a user or module expects real-time termination signals per environment step.
Expected Behavior
self._term_dones or a new parameter should be continuously updated at each compute() call, get_term(name) returns the correct, up-to-date boolean tensor for the current environment step — even when no episode termination occurs.
Suggestion
A minimal fix could be to explicitly clear the term buffer before the loop:
self._term_dones[:] = FalseChecklist
- I have checked that there is no similar issue in the repo (required)
- I have checked that the issue is not in running Isaac Sim itself and is related to the repo