CALVIN MCIL Reproduction (Debug Dataset)

This repo contains my attempt at reproducing results from
CALVIN: A Benchmark for Language-Conditioned Policy Learning.

Setup

• Used the official CALVIN GitHub repo: https://github.com/mees/calvin
• Environment: MacBook Air (CPU only)
• Dataset: Debug split (1.3 GB)

Training Command

See run_command.txt for the exact command used.

Results

• Training ran for 2 epochs on CPU.
• Logs saved in training_logs/metrics.csv.

Notes

• Disabled rollout and tsne callbacks (too heavy for CPU).
• Reduced batch size to 8.
• Limited train batches and validation batches for faster run.