genesis-flowdit-v3-humanoid

FlowDiT V3 Humanoid — Video-to-Navigation (GENESIS)

Part of the GENESIS research framework: video-conditioned robot learning.

Paper: Action Agent: Agentic Video Generation Meets Flow-Constrained Diffusion (IROS 2026)

Code: github.com/jeffrinsam/GENESIS → part2_navigation/flow_constrained_v3_humanoid/

Model Description

FlowDiT V3 Humanoid is an inference-optimized Diffusion Transformer specialized for bipedal humanoid navigation (Unitree G1). It extends FlowDiT V2 with humanoid-specific motion constraints and whole-body balance priors.

Architecture:

• Visual encoder: DINOv2-ViT-B/14 (frozen)
• Flow encoder: RAFT optical flow with humanoid-specific temporal attention
• DiT backbone: Enlarged Diffusion Transformer with balance-constraint cross-attention
• Output: 3-DOF velocity command [vx, vy, yaw_rate] + gait phase signal

Target robot: Unitree G1 humanoid (inference only — see code for the training pipeline).

Runtime: PyTorch 2.9.1+cu128, requires ~4 GB VRAM for inference.

Performance

Evaluated on Unitree G1 in Isaac Sim navigation tasks:

Usage

# Activate the V3 inference venv (torch 2.9.1+cu128)
cd GENESIS/part2_navigation/flow_constrained_v3_humanoid
source .venv/bin/activate

python infer_humanoid.py \
  --checkpoint flowdit_v3_humanoid_best.pt \
  --goal_video goal.mp4 \
  --current_obs obs.jpg

Download via the GENESIS checkpoint script:

bash scripts/download_checkpoints.sh

Checkpoint Details

Citation

@inproceedings{sam2026actionagent,
  title     = {Action Agent: Agentic Video Generation Meets Flow-Constrained Diffusion},
  author    = {Sam, Jeffrin and Khang, Nguyen and Mahmoud, Yara and
               Altamirano Cabrera, Miguel and Tsetserukou, Dzmitry},
  booktitle = {2026 IEEE/RSJ International Conference on Intelligent Robots
               and Systems (IROS)},
  year      = {2026},
  note      = {arXiv:2605.01477}
}

License

Apache 2.0. See LICENSE.