VLA (Vision-Language-Action)

Maps images and language instructions directly to robot actions. Landmarks include RT-2 (web knowledge → control), OpenVLA (open 7B on Open X-Embodiment), and later π0 / RDT / GR00T-style systems. Unlike a VLM, outputs must be executable actions—not text descriptions.

World models

Learn state + action → next observation/state for mental rollouts and planning. Coupled with policies they form WAMs (World-Action Models). Surveys: arXiv:2507.00917, 2510.16732.

Imitation learning / behavior cloning (IL / BC)

Supervised learning from demonstration trajectories. ACT (Action Chunking with Transformers) fits bimanual teleoperation data well—simple and fast to train.

Reinforcement learning (RL)

Optimize policies from rewards, often in massively parallel sim (Isaac Gym, etc.) then transfer via Sim2Real; especially common for locomotion.

Diffusion Policy

Models multimodal action distributions via diffusion (Chi et al., RSS 2023)—strong when demos admit many valid ways to act. Inference needs denoising steps (DDIM etc. help).

Sim2Real

Domain randomization, system identification, Real2Sim2Real, human-in-the-loop correction (e.g., TRANSIC) shrink gaps in physics, rendering, and noise.

Multimodal sensing, tactile & dexterous hands

Vision–language–force–touch fusion; visuo-tactile sensors and e-skins enable fine manipulation. Hardware and algorithms must co-design.

Sources: Embodied-AI-Guide algorithm outline; public ACT/Diffusion Policy comparisons; survey 2407.06886.