---
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.