You can deploy this model and start earning money today!

rlogh/
climbing-holds-rig-no-taxonomy

Name: climbing-holds-rig-no-taxonomy
Availability: InStock

Views

No views yet

climbing-holds-rig-no-taxonomy

DP3-style point-cloud diffusion policy for grasping climbing holds with a Franka arm + LEAP hand, trained on the spring-testbed-era rig dataset (rlogh/climbing-holds-rig).

This is the no-taxonomy ablation. The 64-d grasp-type embedding is removed from the conditioning vector. The model sees only the point cloud + robot state.

The taxonomy-conditioned counterpart lives at rlogh/climbing-holds-rig-with-taxonomy (produced on the other machine).

Files

File	Size	Purpose
`best.pt`	415 MB	EMA weights + optimizer + config dict (best training loss)
`norm_stats.json`	2 KB	Min-max normalization stats — required by `evaluate.py`
`training_status.md`	<1 KB	Final loss + recent-epoch table
`train.log`	~170 KB	Full epoch-by-epoch loss + LR log

Training summary

Dataset: 200 episodes (50 / grasp type), 24,621 valid samples
Epochs: 3000 (cosine LR decay after 500-step warmup)
Batch size: 128, AMP enabled
Best loss: 0.001777 (final 0.001820)
Wall time: 8.5 h on RTX 2080 Ti
Conditioning: PointNet(1024×3) → 256-d + State MLP → 128-d → Fuse → 512-d
U-Net dims: (256, 512, 1024), 100-step cosine DDPM
State / action dim: 23 (Franka 7 + LEAP 16)
--no-grasp-conditioning flag: ON (ablation)

Usage (robot machine)

bash

mkdir -p checkpoints/pc_no_taxonomy_rig
hf download rlogh/climbing-holds-rig-no-taxonomy \
    --local-dir checkpoints/pc_no_taxonomy_rig

python3 data_collection/evaluate.py \
    --checkpoint checkpoints/pc_no_taxonomy_rig/best.pt \
    --pull-dist 0.05 --pull-angle 180 \
    --hold 0 --grasp-type jug

evaluate.py autodetects the encoder type from the embedded config dict. The --grasp-type flag is accepted but ignored at the model level because grasp conditioning is disabled in this checkpoint.

Citation

Repository: github.com/rumilog/rock-climb