RLBench
RLBench copied to clipboard
stepjam
A large-scale benchmark and learning environment.
Currently the gym wrapper resets an episode and simply discards the task descriptions which encode the key information about the target object for a task. ```python def reset(self) -> Dict[str,...
I'm using RLbench for generating a dataset and wanted to use the PourFromCuptoCup task for my work. When I run the code it errors out and says this shouldn't happen....
Is there an easy way to add a little bit of noise to the expert? Like dropping objects and bumping into things (with a small probability) to reach states that...
Hi, thanks for your great work! I was wondering how to extract actions for other action modes from the [imitation learning example](https://github.com/stepjam/RLBench/blob/35c5ae1ff87750e05131cd336da79ce628425567/examples/imitation_learning.py). Specifically for EE_POSE_EE_FRAME. The observation contains the gripper_pose,...
Hello, I'm trying to get demos of the eight tasks used for testing the ARM Algorithm ([https://arxiv.org/abs/2105.14829](https://arxiv.org/abs/2105.14829)) but with a different arm, in this case I changed the 'robot_setup' in...
I was trying to generate open_box task with domain randomization. Episode 1 is fine, but at episode 2, the box is still open (despite it should be closed). This issue...
This benchmark is a very good match for my needs. However, when I use it as an environment for reinforcement learning, I always encounter an action execution failure in the...
Using num_reset attempts correctly for reset_to_demo. Might fix #194 too
Hi, I am trying to generate a multi-task dataset using the provided dataset_generator script. However, when I load the low_dim_obs.pkl, some timesteps seem to miss the _observations.misc["executed_demo_joint_position_action"] property, which makes...
Resolves #204 . There are two callsites for rgb_handles_to_mask, one which reads directly from the sensor (and returns an array on the range [0, 1]) and one which reads from...
