Development blog for TORCS project.
by Lewis
Following up on our development of the parallel training architecture last week, our infrastructure investments have started to pay massive dividends. With multiple headless TORCS instances running in isolated Docker containers, our central PPO algorithm has been absorbing driving experience at an accelerated rate.
However, training a high-performance AI racebot in a vacuum is only half the battle. To ensure our project delivers real business and technical value, our focus this week balanced aggressive model optimization with rigorous preparation for our upcoming stakeholder demonstration with our IBM client.
Blog Overview:
Last week, we noted that reaching a lap time of 1:41.66 was a painfully slow process due to the sample-hungry nature of PPO. This week, our parallelized “pit crew” architecture truly showed its strength.
By scaling up our data collection across parallel network ports, the central AI received a steady, high-velocity stream of driving trajectories. This allowed us to iterate on our reward shaping and hyperparameter tuning much faster than before. The result? Our racebot shattered its previous record, clocking a new fastest lap time of 1:39.13.
The bot is beginning to find tighter racing lines, managing its throttle profiles more aggressively out of corners, and demonstrating a much more sophisticated understanding of the track’s physics limits.
As an AI model evolves, project scopes and technical boundaries naturally shift. To keep our development tightly aligned with corporate expectations, we reviewed and refined our project requirements this week.
We formalized these updated engineering and performance requirements and sent them over to our client. This documentation ensures that both our team and IBM are completely aligned on what constitutes a successful deployment as we head into the final phases of the project
With a record-breaking lap time in hand and a stabilized architecture, we are ready to showcase our progress. We officially reached out to our client via email to schedule a live demonstration meeting.
To make the most of this sync, we are currently building a comprehensive presentation slide deck. This demo will pull back the curtain on the engineering behind our work. The presentation will detail how we bypassed TORCS’ graphical limitations using headless Docker containers, how our orchestration script manages container lifecycles during crashes, and how our parallel framework handles asynchronous weight updates.
We are rapidly approaching a major milestone for this project. Moving from raw execution to client-facing communication requires a shift in focus, but we aren’t letting our technical momentum slow down.
To prepare for our meeting with our client, our immediate action items for the coming week are:
Finalize the demo presentation slides to ensure a polished visual delivery.
Write detailed scripts and speaker notes so the technical presentation is seamless and clear.
Continue pushing model boundaries to see if we can shave even more milliseconds off our 1:39.13 record.