What are the current mainstream and emerging policy architectures in robot imitation learning and VLA, and what are their limitations?

A survey of robot imitation learning and vision-language-action (VLA) models finds that mainstream architectures — ACT-style imitation-trained transformers, Diffusion Policy, and VLAs built on a pretrained VLM backbone plus an action head (RT-2, OpenVLA, π0, π0.5, GR00T, SmolVLA) — remain limited by heavy demonstration requirements, poor performance under distribution shift, and vision-only policies that are blind to contact forces such as mass, friction and stiffness. The overview cites LIBERO benchmarks as near saturated at over 98%, saying they reveal little about real-world generalization, and points to emerging fixes including tactile and force-aware VLAs (Tactile-VLA, TacVLA, HapticVLA, the last of which distills force information so no force sensor is needed at inference), Real-Time Chunking for asynchronous action generation without retraining, and RL post-training on pretrained VLAs for reliability and data efficiency.

Welcome Short version, based on recent surveys and papers: 1. Mainstream - Chunked policies trained by imitation: ACT-style transformers and Diffusion Policy. - VLAs: a pretrained VLM backbone plus an action head. Two families: autoregressive action tokens RT-2, OpenVLA and diffusion/flow-matching heads π0, π0.5, GR00T, SmolVLA . Dual-system designs a slow VLM planner plus a fast low-level policy are common too. See this survey https://arxiv.org/abs/2509.19012 . 2. Limitations - Data: VLAs need a lot of demos, and performance drops under distribution shift. - Physical properties: vision doesn’t show mass, friction or stiffness, so a vision-only VLA is blind to contact forces https://arxiv.org/pdf/2606.29089 . - Latency: big models are slow, so actions can be computed from stale observations. - Evaluation: benchmarks like LIBERO are near saturated 98% , so they say little about real generalization survey https://www.preprints.org/manuscript/202604.0664 . 3. Emerging directions matching your interests - Tactile/force VLAs for weight, friction and stiffness: Tactile-VLA https://arxiv.org/abs/2507.09160 , TacVLA https://arxiv.org/abs/2603.12665 , and HapticVLA https://arxiv.org/abs/2603.15257 , which distills force info so no force sensor is needed at inference. - Real-time execution: Real-Time Chunking asynchronous chunk generation, works with flow-based VLAs without retraining , and RL fine-tuning for real-time control Real-Time EXPO-FT https://alphaxiv.org/abs/2609.18207 . - RL post-training on top of pretrained VLAs for reliability and data efficiency. Tip: you also posted this in Research, so keeping one thread will help you get answers in one place.