Repository navigation
The DHWANI rover as it stood after the SIH26126 idea submission (1 October 2026).
Mission parser weights. These files are too large for git. Save each one in its model folder as weights.npz:
| File | Save as | Size | MD5 |
|---|---|---|---|
tinybert4-weights.npz |
parser/models/tinybert4/weights.npz |
56.5 MB | 92a40887b235c3ddd24826b246932143 |
berttiny-weights.npz |
parser/models/berttiny/weights.npz |
17.3 MB | b031465b8c794129c6cc41591963be4a |
TinyBERT-4L (the one the rover uses) gets 97.75 % of the 2,000 validation instructions exactly right and 98.4 % of 192 written by hand. On the Jetson Nano's CPU its median parse time is 45 ms with the rover parked, using the NumPy 1.19 OpenBLAS build described in docs/SETUP.md.
Then check them with python3 parser/rover_check.py on the rover, or python3 parser/parity_test.py && python3 parser/evaluate.py on a laptop.