feat: add track completion detection and comprehensive environment guide

[EldarAlvik]

Thank you for the project and wanted to contribute a bit with some extra documentation that i have gotten from working on the gym environment using the get_environment method and added info extraction about if the agent has reached the finish line. also a tiny semantics thing for compute reward.

Changes

1. Add reached_finishline flag (tmrl/custom/tm/tm_gym_interfaces.py)

• Explicitly indicates successful track completion in info dict
• Solves ambiguity: episodes end for multiple reasons (success, timeout, off-course)
• Enables getting a rough time for completteion
• Backward compatible

2. Create gymnasium environment documentation (readme/tmrl_gym_enviroment.md)

• Moved existing documentation from README to dedicated guide
• More in-depth explanation of how to setup and use TMRL as purely a Gym environment

3. Link gymnasium environment guide in README (README.md)

**4. Updated terminated/truncated semantics slightly for tmrl/custom/tm/utils/compute_reward.py

• Rename terminated → truncated for timeout/off-track failures
• Follows Gymnasium spec: terminated = goal state, truncated = constraint violation
• Only crossing finish line is true termination

Testing

• Tested with python -m tmrl --test
• Tested with TMRL standard running workflow with server, trainer and worker.
• Tested with custom agent using get_environment().

[yannbouteiller]

Nice, thanks for the contribution, I will try to look into it soon.

I am unsure about what you are describing regarding the distinction between termination and truncation, though. The way I understand it, termination means MDP terminal state (which includes going off-limits) whereas truncation means episode truncated in a non-terminal state (i.e., ending the episode due to some non-observable constraint, typically time limit - the point is that an episode should not be considered terminated in a truncated state, i.e., the value function should not be considered 0 at that time, as the constraint is not part of the MDP ; the off-limits constraint is clearly a terminal state: the entire point of this constraint is to set the value function to 0 when the car strays too far from the road)

[EldarAlvik]

ok. i guess it could be argued either way as the agent could continue acting. But after looking at how car_racing handles it i agree with you. I updated the pull request and undid the change and updated the readme