Skip to content

Wrong teacher normalizer input size in StudentTeacherRecurrent #115

New issue
New issue
Open
#116
Open
Wrong teacher normalizer input size in StudentTeacherRecurrent#115
#116
@tianzong-cheng

Description

@tianzong-cheng
tianzong-cheng
opened on Sep 20, 2025

The teacher normalizer is initialized with rnn_hidden_dim, which I believe is incorrect according to evaluate()

rsl_rl/rsl_rl/modules/student_teacher_recurrent.py

Lines 80 to 93 in cf71aa6

# teacher
if self.teacher_recurrent:
self.memory_t = Memory(
num_teacher_obs, type=rnn_type, num_layers=rnn_num_layers, hidden_size=rnn_hidden_dim
)
num_teacher_obs = rnn_hidden_dim
self.teacher = MLP(num_teacher_obs, num_actions, teacher_hidden_dims, activation)
# teacher observation normalization
self.teacher_obs_normalization = teacher_obs_normalization
if teacher_obs_normalization:
self.teacher_obs_normalizer = EmpiricalNormalization(num_teacher_obs)
else:
self.teacher_obs_normalizer = torch.nn.Identity()

rsl_rl/rsl_rl/modules/student_teacher_recurrent.py

Lines 161 to 168 in cf71aa6

def evaluate(self, obs):
obs = self.get_teacher_obs(obs)
obs = self.teacher_obs_normalizer(obs)
with torch.no_grad():
if self.teacher_recurrent:
self.memory_t.eval()
obs = self.memory_t(obs).squeeze(0)
return self.teacher(obs)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions