BbEeNn1314/robosyn-challenge-weights

RoboSyn Challenge Submission - Hybrid ACT+PP Policy

Competition: NeurIPS 2026 RoboSynChallenge
Team: YongbinChen
Average Success Rate: 92.7% (across 10 manipulation tasks)

Model Description

This submission uses a hybrid approach combining:

• ACT (Action Chunking Transformer) for 3 tasks requiring learned visuomotor skills
• Perception-Planning (PP) pipelines for 7 tasks amenable to geometric reasoning

Architecture

ACT Tasks (Neural Network):

• table_rearrangement: 95.7% success rate
• click_bell: 99.3% success rate
• item_assembly: 81.2% success rate

PP Tasks (Rule-based Perception):

• sample_loading: 88.4% success rate
• items_handover: 87.0% success rate
• manipulate_pipette: 90.6% success rate
• drawer_open_place: 95.7% success rate
• handle_basket: 96.4% success rate
• water_pouring: 96.4% success rate
• mixer_operating: 96.4% success rate

Training Details

ACT Models:

• Dataset: Official RoboSynChallenge demonstrations (1000 episodes per task)
• Framework: LeRobot 0.4.4
• Backbone: ResNet-18 (3 camera views: wrist, front, side)
• Training: 40k-80k steps, batch size 32
• Hardware: 2× RTX 4090 24GB

PP Pipelines:

• Zero-shot perception using Open3D + OpenCV
• Hand-crafted motion primitives
• Iterative refinement strategies

Usage

See the code repository for:

• Installation instructions
• Evaluation scripts
• Per-task configuration

# Install dependencies
uv sync

# Run evaluation (46 episodes per task)
bash scripts/eval_all_tasks.sh

Performance Metrics

Task	Method	Success Rate	Avg Steps
table_rearrangement	ACT	95.7%	302
click_bell	ACT	99.3%	180
item_assembly	ACT	81.2%	410
sample_loading	PP	88.4%	285
items_handover	PP	87.0%	303
manipulate_pipette	PP	90.6%	349
drawer_open_place	PP	95.7%	245
handle_basket	PP	96.4%	198
water_pouring	PP	96.4%	220
mixer_operating	PP	96.4%	215
Average		92.7%	271

Files

This repository contains trained checkpoints for the 3 ACT models:

• click_bell/pretrained_model/ - Click bell task weights
• item_assembly/pretrained_model/ - Item assembly task weights
• table_rearrangement/pretrained_model/ - Table rearrangement task weights

PP task configurations are in the code repository (no trained weights required).

Citation

@misc{robosyn2026submission,
  author = {YongbinChen},
  title = {RoboSyn Challenge Submission - Hybrid ACT+PP Policy},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/BbEeNn1314/robosyn-challenge-weights}}
}

License

Apache 2.0