Add SwipeActionWrapper for gesture-level swipe actions

[kavyabhand]

Summary

Adds SwipeActionWrapper, a higher-level action wrapper that converts a swipe (start → end position) into a sequence of interpolated TOUCH steps followed by a LIFT at the end position.

This complements the existing TapActionWrapper and supports the documented use case of hard-coding gesture skills for RL studies where agents need swipe/scroll primitives without manually chaining raw touch/lift actions.

Changes

• android_env/wrappers/swipe_action_wrapper.py: new wrapper with:
  • Action spec: start_position and end_position (BoundedArray shape (2,) in [0, 1])
  • Configurable num_steps (default 10) for interpolation granularity
  • Reward accumulation across sub-steps
  • Early return on StepType.LAST (same behavior as TapActionWrapper)
  • env_steps tracking in stats()
• android_env/wrappers/swipe_action_wrapper_test.py: unit tests for interpolation, single-step edge case, reward accumulation, early termination, and specs

Example usage

from android_env.wrappers import swipe_action_wrapper
import numpy as np

env = swipe_action_wrapper.SwipeActionWrapper(env, num_steps=10)
env.step({
    'start_position': np.array([0.5, 0.8], dtype=np.float32),
    'end_position': np.array([0.5, 0.2], dtype=np.float32),
})

[kenjitoyama]

Can we simplify this with np.linspace() or something like that?

[kenjitoyama]

I believe the total reward should be discounted, otherwise rewards coming later would have the same value as the earlier rewards.

[kenjitoyama]

Hi, thanks for submitting this. I thought we had this wrapper, but somehow we don't. I'm pretty sure we used something just like this in https://arxiv.org/abs/2204.10374, but I guess we forgot to open source it.

This is the first time that I'm trying to merge a PR into Google's internal AndroidEnv version, so we might face a few hiccups. Hopefully it'll be fine.

[kavyabhand]

@kenjitoyama
Thank you so much for reviewing! I have tried implementing the changes suggested by you, please if you could review those and let me know if there is anything else I could contribute you.

tldr;

1. Replaced the manual interpolation loop with np.linspace (also removes the num_steps == 1 special case).
2. Sub-step rewards are now accumulated with discounting (reward_discount *= discount per sub-step), with a test for the discounted case.