An environment-aware prediction and control framework, which incorporates learned environment and terrain features into a predictive model for human-robot symbiotic walking, is disclosed herein. First, a compact deep neural network is introduced for accurate and efficient prediction of pixel-level depth maps from RGB inputs. In turn, this methodology reduces the size, weight, and cost of the necessary hardware, while adding key features such as close-range sensing, filtering, and temporal consistency. In combination with human kinematics data and demonstrated walking gaits, the extracted visual features of the environment are used to learn a probabilistic model coupling perceptions to optimal actions. The resulting data-driven controllers. Bayesian Interaction Primitives, can be used to infer in real-time optimal control actions for a lower-limb prosthesis. The inferred actions naturally take the current state of the environment and the user into account during walking.