Poster: Zero-shot Safety Prediction for Autonomous Robots with Foundation World Models
0 likes55 views
The document discusses a novel approach to safety prediction for autonomous robots using a training-free world model that integrates foundation models. It addresses challenges such as predicting observations at the object level and out-of-distribution prediction shifts, while providing metrics for evaluating learned dynamics and safety assessments. Key contributions include the development of a segmentation-based metric for dynamics and an interpretable representation for direct safety evaluation.
Poster: Zero-shot Safety Prediction for Autonomous Robots with Foundation World Models
- 1. ■ Low-quality observation prediction by world model W given current observation yt and action ut ■ Fine-tuning of foundation models ■ Additional dynamical states like speed and angular velocity ■ Apply multimodal foundation models ■ Case studies on physical robots ■ Unknown states and dynamical models Preprint: Mao, Z.; Dai, S.; Geng, Y and Ruchkin, I. "Zero-shot Safety Prediction for Autonomous Robots with Foundation World Models" arXiv (2024). Zhenjiang Mao, Siqi Dai, Yuang Geng, Ivan Ruchkin Trustworthy Engineered Autonomy (TEA) Lab Department of Electrical and Computer Engineering, University of Florida Zero-shot Safety Prediction for Autonomous Robots with Foundation World Models APPROACH FUTURE WORK RESULTS: LUNAR LANDER CHALLENGES & CONTRIBUTIONS PROBLEM REFERENCES ─ yt+1 = W (yt ,ut ) ≠ yt+1 ^ ■ How to predict/evaluate the observation at object level? ─ MSE( yt+1 ,yt+1 ) ^ Challenges: Contributions: ■ A training-free world model combines foundation models and overcomes distribution shift in existing world models ■ A segmentation-based metric to measure the learned surrogate dynamics by comparing object-level error ■ An interpretable latent representation that can evaluate safety directly and improve safety predictions ■ Need to train specialized safety evaluator NN: y→ {T, F} ■ Out-of-distribution (OOD) shift in predictions World models: Ha, D.; and Schmidhuber, J. 2018. Recurrent World Models Facilitate Policy Evolution. In Advances in Neural Information Processing Systems, volume 31. Safety prediction: Mao, Z.; Sobolewski, C.; and Ruchkin, I. "How Safe Am I Given What I See? Calibrated Prediction of Safety Chances for Image-Controlled Autonomy." In Learning for Dynamics and Control Conference 2024. Image Rebuilding: Move the objects from previous position wt into the predicted position wt+1 Coarse evaluation: Object duplication Object loss Ground truth Ground truth ─ Object Centroid Distance: CD = ||( w i t+1 )C -( w i t+1 )C || Fine-grained evaluation: Observation: y Segmented objects: w 1 -w 4 ^ F1 score and False Positive Rate (FPR) of safety prediction SSIM and MSE of predicted observation 10 20 30 40 50 60 Prediction Horizon 10 20 30 40 50 60 Prediction Horizon 10 20 30 40 50 60 Prediction Horizon 10 20 30 40 50 60 Prediction Horizon 10 20 30 40 50 60 Prediction Horizon 10 20 30 40 50 60 Prediction Horizon 1.0 0.8 0.6 0.4 0.2 0.0 Standard World Model Segmentation + Supervised World Model Foundation World Model VAE + MD-LSTM SAM + LSTM SAM + GPT3.5 SAM + Gemma Dashed lines indicate the use of supervised training data. 1.00 0.95 0.90 0.85 0.80 0.75 0.70 0.200 0.175 0.150 0.125 0.100 0.075 0.050 0.025 0.000 5 4 3 2 1 0 7 6 5 4 3 2 1 0 Before After Before After ^ 1.0 0.8 0.6 0.4 0.2 0.0 Horizontal and vertical position error