File size: 1,669 Bytes
19a2511 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 | ---
library_name: pytorch
tags:
- robotics
- actuator-model
- gru
- g1
- h1
---
# AnchorLab Models
Presentation actuator-model checkpoints for the AnchorLab G1 and H1 right-arm
benchmarks. Each ZIP contains the definitive TorchScript artifact, raw training
checkpoint, normalization statistics, configuration files, a smoke test, an
I/O example, and detailed deployment notes.
## Models
### G1 presentation full-torque GRU
File: `g1_presentation_full_torque_gru.zip`
This is the exact seed-48 G1 checkpoint used for the presentation result. It is
the original 24-input model, not the later 20-input ablation.
Deployment contract:
- Full-torque GRU for the fixed-pelvis G1 29-DOF tri-hand benchmark.
- Solver PD must be disabled: `kp=0`, `kd=0`.
- The 24 inputs are `q`, position error, velocity, `pd_hint`, `qfrc_bias`, and
previous torque.
- The GRU output is the complete applied motor torque.
### H1 presentation residual GRU
File: `h1_presentation_residual_gru.zip`
This is the exact H1 right-arm v28 checkpoint used for the presentation result.
Deployment contract:
- Residual GRU for the measured fixed-root H1 gripper/end-effector setup.
- Run as implicit SysID PD plus residual:
`applied_torque = implicit_sysid_pd_torque + 0.33 * residual`.
- Do not run this checkpoint as a pure full-torque model with `kp=0` and
`kd=0`.
- The 24 inputs are `q`, position error, velocity, `pd_hint`, `qfrc_bias`, and
previous torque.
## Scope
These archives are actuator-model handoffs, not complete IsaacLab-Newton
checkouts. Map the included reference configurations and asset paths into the
benchmark checkout where the model is evaluated.
|