DEV Community

kiwi_tech
kiwi_tech

Posted on • Originally published at kiwi-tech.hashnode.dev

Kiwi-chan's Rocky Road to Resourcefulness! 🥝⛏️

Kiwi-chan's Current View

Hey everyone! Another four hours down with Kiwi-chan, our autonomous Minecraft AI, and it's been... a journey. We're seeing steady progress, even if it doesn't always look like it! It's like watching a toddler learn to walk – lots of stumbles, but eventually, they get there.

We've been focusing on expanding Kiwi-chan's skill library, and she's definitely trying! We've added (and re-attempted!) skills like explore_forward, craft_stone_pickaxe, goto_crafting_table, and gather_dirt. The logs show a lot of back-and-forth, especially around getting to and using that crafting table. It seems pathfinding is still a bit of a challenge in certain terrains.

📊 Current Progression & Stats

Over the last four hours, Kiwi-chan attempted a total of 110 tasks. We had a 50.0% success rate – 55 successes and 55 failures. While a 50/50 split might not sound amazing, remember this is learning! Each failure is a data point, a lesson learned, and a step closer to a more robust and reliable AI. We're seeing a lot of repetition in the failures – specifically around crafting a stone pickaxe and getting to the crafting table – which is incredibly valuable for debugging and refining the system. The AI is also getting "bored" of repeating tasks, which is a good sign that it's recognizing unproductive loops!

I've been working with Qwen to provide recovery plans when things go south, and it's been a huge help in steering Kiwi-chan back on track. We're also constantly refining the core rules to prevent these loops and encourage more efficient exploration. The logs show a lot of "Fixing Code" attempts, and I'm happy to report that some of those fixes are sticking!

It's fascinating to see how Kiwi-chan responds to different situations. The "oak obsession ban" is working well, forcing her to consider alternative resources when oak logs aren't available. The emphasis on dynamic block finding (no hardcoded coordinates!) is also paying off.

The biggest challenge right now seems to be navigating the environment and reliably interacting with blocks. We're seeing a lot of "Failed to move. Blocked by obstacle." errors, which suggests we need to improve the pathfinding and obstacle avoidance algorithms.

Overall, it's a slow and steady climb, but I'm incredibly excited about the progress we're making. Kiwi-chan is learning, adapting, and becoming more resourceful with each passing hour.

🥝✨ Help Kiwi-chan Level Up! ✨🥝

This project is a labor of love, and it's running on a… let's just say enthusiastically cooled Frankenstein rig. Any support you can offer will directly help me upgrade to an RTX 3060, which will significantly speed up training and allow us to tackle even more complex challenges!

Top comments (0)