mcMotionChallengeGitHub ↗

MODEL / CHECKPOINT

Amazon FAR

Model

Author TWIST2 1017-20k network evaluated with the sim2real deployment configuration: 50 Hz, first-frame history, wrist Kp 40 / Kd 5 and unclipped actions.

Organization
Amazon FAR
Type
Dense
Convolutional history encoder and dense policy layers.
Version
1017 · 20k
author checkpoint filename
Robot support
Unitree G1 · 29 DoF
Author source revision
b06178f19a22f2138cbd31f60c6d494bc263f67d

Checkpoint artifacts

Author release
amazon-far/TWIST2 ↗
Artifact type
Author-released checkpoint
Original file
twist2_1017_20k.onnx ↗

The adapter pins this published artifact by SHA256. Artifact identity is separate from inference parity.

Download links deliver the adapter files. The release links and verification notes distinguish author files from conversions; SHA256 identifies the bytes used.

Policy checkpoint
Download ↗
2d1fb3a31e4e967f70ecfefc3ad1e7b2ac491677068b89f60b565a94e7735061

Inference interface

The adapter declares its control clock and reference window below. Source and configuration describe the observation and action transforms.

Observation requirements

Observation rules ↗

Model-specific handling

Original TWIST2 1017-20k network with the pinned sim2real deployment: 50 Hz, first-observation history fill, wrist Kp 40/Kd 5, and unclipped network actions. The author sim2sim route instead uses 100 Hz, zero history, wrist Kp 4/Kd 0.2 and action clipping at +/-10. Benchmark reference preparation, reset, 2 ms physics and scoring stay in Core.

Exact checkpoint & evidence ↗
Control interval
20 ms
Reference history
0 ms
Reference preview
0 ms
Output interface
joint-pd-feedforward
Configuration
policy.json

Evaluations