LaMP: Learning Vision-Language-Action Policies with 3D Scene Flow as Latent Motion Prior

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• LaMP consistently outperforms evaluated VLA baselines across LIBERO, LIBERO-Plus, and SimplerEnv-WidowX benchmarks, achieving the highest reported average success rates under the same training budgets.result 90%
• On LIBERO-Plus OOD perturbations, LaMP shows improved robustness with an average 9.7% gain over the strongest prior baseline.result 90%
• LaMP addresses the limitation of existing VLA models that degrade under unfamiliar spatial dynamics by embedding dense 3D scene flow as a latent motion prior.technical 85%
• The approach uses substantial computational resources for training, and the requirement for 3D scene flow understanding may limit its applicability in environments with certain resource constraints.limitation 80%
• LaMP aligns a flow-matching Motion Expert with a policy-predicting Action Expert using gated cross-attention.method 95%
• The Motion Expert generates a one-step partially denoised 3D scene flow, and its hidden states condition the Action Expert without full multi-step reconstruction.method 85%
• Existing VLA models regress actions directly from 2D semantic visual features, forcing them to learn complex 3D physical interactions implicitly.technical 80%
• LaMP can potentially replace less advanced robotic control systems that rely heavily on 2D inputs, offering more robust and adaptable solutions for complex manipulation tasks.market 70%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card