ravediamond/act-aloha-sim-transfer-cube-gguf

ACT, ALOHA sim transfer cube, GGUF for vla.cpp

lerobot/act_aloha_sim_transfer_cube_human converted to GGUF for vla.cpp. Weights are unchanged F32. The ResNet's frozen batch norms are folded into the convs, the VAE encoder is left out (the latent is zero at inference), and the normalizer stats are stored in the file, so vla.cpp takes raw state and pixels and returns actions in robot units.

One top camera at 480x640, 14 dim state and action, 100 action chunk.

Run

./build/vla-server act-aloha-sim-transfer-cube-f32.gguf
./build/vla-bench --ckpt act-aloha-sim-transfer-cube-f32.gguf --images 1 --size 640 --height 480

On a Jetson Orin NX 16 GB one chunk takes 25.6 ms (22.3 ms with --weight-dtype f16 --flash-attn).

Parity

Against LeRobot's PyTorch ACTPolicy on the same synthetic 640x480 input, the max abs action error is 2e-7 on CPU and 2e-4 on CUDA and Metal, on actions that span about 1.4.

Convert it yourself

python scripts/convert_act_to_gguf.py --ckpt <path to the downloaded checkpoint> --out act-aloha-sim-transfer-cube-f32.gguf

The original model is Apache 2.0, by the LeRobot team. ACT is from Zhao et al. 2023.

Write-up and register

• Porting ACT to vla.cpp
• Every paper re-tested on this rig, with its bar, verdict and code: the register