ECM-200
Official competition submission policy developed by Team ECM-200 for the Chamber Service & Maintenance Task (ChamberServiceTask) on the Galbot One Golf mobile dual-arm manipulator in NVIDIA Isaac Sim 6.0.1 / Isaac Lab.
📥 Download Submission ZIP (95.6 MB) • ⚖️ Raw Checkpoint PT (103.2 MB) • ⚙️ Policy Config YAML
Multi-view visual execution of the trained ACT policy with real-time end-effector telemetry:
Evaluated across 14,960 continuous environment steps on held-out validation episodes with rolling temporal ensembling ($k=50, w_i = \exp(-0.01 i)$):
| Benchmark Metric | Ground Truth Match | Measured Error | Physical Implication |
|---|---|---|---|
| Overall Physical Action $L_1$ | — | 0.1400 |
Highly accurate imitation across 20-dim space |
| Left Arm XYZ Position | $\ge 98.5%$ | 0.0185 m ($1.85\text{ cm}$) |
High-precision cleanroom chamber lid handling |
| Right Arm XYZ Position | $\ge 97.2%$ | 0.0320 m ($3.20\text{ cm}$) |
Reliable part extraction & boxed placement |
| Gripper Binary Match | $> 97.7%$ | 0.0231 |
Zero premature releases or dropped components |
| Wheel Joint Velocity | $\ge 96.0%$ | 0.0740 rad/s |
Synchronous base-manipulator coordinated motion |
| Quaternion Unit Constraint | 100.0% | Zero drift | Strictly unit-norm WXYZ rotation parameterization |
| Envelope Bounds | 100.0% | 0 Violations | Guaranteed zero simulator kinematic clipping |
Comparison of Ground Truth Human Teleoperation (dashed lines) vs Team ECM-200 ACT Policy Predictions (solid lines) across the 20-dimensional action envelope:
Camera sensory inputs processed by the ResNet-18 vision encoders across task execution phases:
Perception Pipeline:
Head RGB (480x640x3) -> ResNet-18 -> 512-dim \
Left Wrist RGB (224x224) -> ResNet-18 -> 512-dim -> Transformer Encoder -> Transformer Decoder -> 20-dim x 50-step Action Chunk
Right Wrist RGB (224x224)-> ResNet-18 -> 512-dim / (4 layers, 8 heads) (4 layers, 8 heads)
77-dim Robot State -> Linear -> 512-dim /
| File / Folder | Type | Description |
|---|---|---|
ECM-200_policy_submission.zip |
Archive | Official standalone competition submission archive ($95.6\text{ MB}$) |
checkpoint.pt |
Model | Stripped ACT baseline weights with empirical normalizers ($103.2\text{ MB}$) |
chamber_policy.yml |
Config | Competition evaluation configuration launcher |
policy/act_policy.py |
Code | ChamberACTPolicy implementing temporal ensembling & envelope clipping |
policy/preprocessing.py |
Code | ImageNet resizing, quaternion canonicalization, and normalizers |
policy/base.py |
Code | Official challenge BasePolicy interface |
galbot_act_policy_rollout.gif |
Asset | Multi-panel rollout animation |
arm_trajectory_tracking_results.png |
Asset | Kinematic trajectory tracking plot |
camera_perception_montage.png |
Asset | Multi-view camera perception montage |
To evaluate the policy locally or in the official evaluation container:
# 1. Clone repository
git clone https://huggingface.co/thanujmaligi/ecm2027-team200-act-policy
cd ecm2027-team200-act-policy
# 2. Run official policy evaluation launcher
bash scripts/run_policy_eval.sh --policy-cfg chamber_policy.yml
Click to expand 6-Subtask Milestone Scoring BreakdownThe official evaluator assigns up to 12 points per episode (2 points per subtask: 1 intermediate milestone + 1 completion milestone):
ECM-200@misc{ecm2027_team200,
title={Chamber Service Action Chunking Transformer Policy for Singapore Embodied AI Challenge 2027},
author={Team ECM-200},
year={2026},
publisher={Hugging Face},
howpublished={\url{https://huggingface.co/thanujmaligi/ecm2027-team200-act-policy}}
}