Abstract: Disclosed herein are systems and methods for quantizing floating-point values during operation of robotic systems. For example, a system can be configured to receive input data having multiple values represented in a floating-point format and determine a maximum absolute value from these values. The system can adjust the representation of each value by removing at least one bit from the exponent portion and redistributing it to the mantissa portion. In examples, the system can also extract a sign bit and align each value by shifting each bit. Based on the adjusted values, the system can perform dynamic quantization of the input data based at least in part on a scaling value determined using the maximum absolute value. The input data can then be provided as an input to at least a portion of a neural network during execution of one or more operations by the neural network.

Abstract: Disclosed herein are methods and systems for a navigation paradigm. One method includes receiving a set of quantized values resulting from a matrix multiplication of quantized query and key vectors, the set of quantized values corresponding to sensor data associated with the ego; identifying a maximum value using an integer comparison operation; for at least one quantized value in the set of quantized values, computing a difference between the quantized value and the maximum value; executing a bit shift operation, on the difference between the quantized value and the maximum value, corresponding to a power-of-two quantization scale to generate bit-shifted values; executing a base-2 exponentiation operation to the bit shifted values to generate base-2 exponentiated values; aggregating the base-2 exponentiated values to generate an aggregated value; and executing a model using the aggregated value to output a navigational instruction to be executed by the ego.

Abstract: Embodiments described herein provide for computing masked scores in transformer models without materializing a full mask. A matrix is partitioned into submatrices based on available processing elements. At least one element is initialized with row and column index values and evaluates a masking condition using a comparison operation, which is fused into the softmax computation. This approach supports causal and batched masks with minimal overhead and is well-suited for applications such as autonomous navigation, where computed weights guide downstream decision-making.

Abstract: Embodiments include systems and methods for processing sensor data and generating operational instructions for automated devices. The automated device includes hardware (e.g., INT8 MAC units, FP32 SIMD processors) and software operations executing a machine-learning architecture having transformer architectures implementing mixed-precision quantization for ingesting and analyzing various types of sensor data. The machine-learning architecture implements a transformer that includes dynamic quantization layers that dynamically quantize tensors, per-token, for downstream operations of the transformer and machine-learning architecture.

Abstract: Embodiments include systems and methods for processing sensor data and generating operational instructions for automated devices. The automated device includes hardware (e.g., INT8 MAC units, FP32 SIMD processors) and software operations executing a machine-learning architecture having transformer architectures implementing mixed-precision quantization for ingesting and analyzing various types of sensor data. The machine-learning architecture implements a transformer that includes dynamic quantization layers that dynamically quantize tensors, per-token, for downstream operations of the transformer and machine-learning architecture.

Abstract: Embodiments include systems and methods for rotary positional embedding in a mixed-precision pipeline. An example method can be performed in a mixed-precision pipeline including a multiplier-accumulator (MAC) for a first bit width to execute a multiplication function and a logic execution block for a second, greater, bit width. The method includes operations executed by the circuit including obtaining, via a circuit for a first bit-width, an input tensor and a logarithm of an angle, ?; generating, by a multiplication function for inputs having the first bit-width, a product of a first element of the input tensor and a first element of the logarithm of ?, each of the first elements having the first bit-width. The method includes operations executed by the logic execution block including generating an exponent of the product to determine ? according to the second bit-width and generating a rotation matrix according to trigonometric functions of ?.

Abstract: Embodiments include systems and methods for input of raw Bayer image input. The method includes obtaining input data including multiple data elements having a second bit-width exceeding the first bit-width. The method includes generating multiple sets of the input data, each set having a portion of the data elements of the input data, by convolving a various first predefined kernels with the input data, each of the various first predefined kernels corresponding to one of the sets. The method includes, for each set, generating multiple subsets of the set by convolving a plurality of second predefined kernels with the set, each of the second predefined kernels corresponding to one of the subsets. The method includes generating channels, each of which includes output data including one or more of the multiple subsets. The method can be performed by a circuit for a first bit width.

Abstract: Embodiments include systems and methods for special input formatting for binary input data to be able to process the binary input data as smaller binary data. A method can be performed by a circuit comprising a multiplier-accumulator (MAC) for convolving a data structure with weights of a machine learning model. The method includes obtaining first and second values having a first bit-width. The method includes generating, using a multiplication function of the MAC, a first product using the first value and a first predefined weight, the multiplication function, thereby left shifting the first product. The method includes generating, using the multiplication function, an output word using the first product, the second value, and a second predefined weight. The method includes storing, upon receipt of an instruction, a respective first and second portion of the output word to a first and second addressable location.

Abstract: Embodiments include systems and methods for bit-augmented computation. A method can be performed by a circuit for a first bit-width. The method includes obtaining an input data structure including multiple elements of a second bit-width, greater than the first bit-width. The method includes generating a first and a second data structure from the input data structure, the first data structure and the second data structure having a bit-width which does not exceed the first bit-width. The method includes executing, by the circuit, one or more layers of a neural network of a machine-learning architecture to generate a first output, the one or more layers of the neural network taking as inputs the first data structure and a set of one or more weights.