Modelpublic

knightnemo/robotwin-icl-v3-arx-x5-vam-ti2v5b-openwam-ref-100k

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

0likes

Model Card

RoboTwin ICL v3 ARX-X5 VAM TI2V-5B OpenWAM Ref 100K

This repository contains a 100k-step VAM checkpoint trained on the RoboTwin ICL-paired v3 dataset for the ARX-X5 target robot.

This is the reference-video variant: during training the model used cross-embodiment paired reference videos from the same task/seed when available. It is intended for ICL-style human/robot reference conditioning, current target-view anchoring, and action rollout prediction.

Files

text

step-100000.safetensors   # joint VAM checkpoint at 100k steps
model_config.json         # architecture and data/preprocessing contract
training_log_node0.txt    # rank/node 0 training log
training_log_node1.txt    # rank/node 1 training log
README.md                 # this model card

The checkpoint contains both fine-tuned Wan video DiT weights and action_dit.* ActionMoT weights. The Wan2.2-TI2V-5B base assets are not bundled.

Training Summary

text

run directory:    src/vam/models/train/vam_icl/paired_v3_alltasks_mv_mot_ti2v5b_16g_100k_v2_ref_openwam_fastwam_warmstart
checkpoint:       step-100000.safetensors
dataset:          RoboTwin ICL-paired v3
target robot:     arx-x5
task split:       20 train tasks
episodes/task:    150 train episodes
backbone:         Wan2.2-TI2V-5B
action stream:    30-layer ActionMoT, OpenWAM/FastWAM warm-start style
reference mode:   enabled, paired cross-embodiment reference video
multiview:        enabled, RoboTwin head + left wrist + right wrist layout

Action And Proprio Contract

text

action_space:     joint
proprio_space:    joint
action_dim:       14
proprio_dim:      14
action_format:    absolute

The 14-D action/state vector follows RoboTwin joint_action/vector:

text

left arm joints + left gripper + right arm joints + right gripper

The model predicts absolute joint targets, not delta actions.

Temporal And Visual Contract

text

raw action window:          33 frames
action horizon:             32 actions
action_video_freq_ratio:    4
target video frames:        9 frames sampled at raw indices [0, 4, ..., 32]
resolution:                 384 x 320
multiview layout:           head camera on top, left/right wrist cameras below
full reference video:       enabled
max reference frames:       41
reference subsample factor: 4

For closed-loop eval, use:

text

action_dim=14
proprio_dim=14
num_frames=9
action_horizon=32
height=384
width=320
multiview=true
action_space=joint
proprio_space=action

Notes

This checkpoint was trained for research use in the local VAM codebase. It expects the matching model architecture, action normalization stats, and RoboTwin observation/action conventions used by the training scripts in src/vam/examples/wanvideo/human2robot.