Reference
Glossary
Short, precise definitions of the terms used across the courses. Each entry links to the lesson that covers it in depth.
A
- Action chunking
- Predicting a sequence of future actions at each step instead of a single action, which makes learned policies smoother and reduces compounding errors. It is central to the ACT policy.
- Application-specific integrated circuitASIC
- A chip designed and fabricated for one function. An ASIC has no programmable fabric, so it is smaller, faster and lower power per function than an FPGA, but each design change requires a new fabrication run with high fixed costs.
- Asynchronous FIFO
- A FIFO whose write and read sides run on unrelated clocks. Its read and write pointers cross between domains as Gray codes through synchronisers, which makes it the standard way to move data streams between clock domains.
- AXI4-Lite
- A simple member of Arm's AMBA AXI bus family for register-style access: single transfers on separate read and write channels, each using a VALID/READY handshake.
B
- Behaviour cloning
- Learning a robot policy by supervised learning on recorded demonstrations, mapping observations to the actions the demonstrator took.
- Bitstream
- The binary configuration file that programs an FPGA. It sets every LUT, routing switch, memory initial value and I/O setting. Vendor tools generate it at the end of implementation.
- Block RAMBRAM
- Dedicated on-chip memory blocks in an FPGA, typically a few kilobits to tens of kilobits each, with one or two synchronous ports. Synthesis tools infer block RAM from arrays that are read and written with registered, clocked access.
- Blocking and non-blocking assignment
- In Verilog, a blocking assignment (=) updates its target immediately, while a non-blocking assignment (<=) schedules the update for the end of the time step. Use non-blocking assignments for registers in clocked blocks and blocking assignments in combinational blocks.
C
- Camera calibration
- Estimating a camera's intrinsic parameters and lens distortion, usually by photographing a known pattern such as a checkerboard from several views and solving an optimisation.
- Camera intrinsics
- The parameters of a camera's internal geometry, focal lengths and principal point, collected in the matrix K, which maps points in the camera frame to pixel coordinates. Lens distortion is modelled separately.
- Carry chain
- Dedicated fast routing and logic between adjacent logic cells that propagates the carry of an adder or counter. Synthesis tools map arithmetic such as a + b onto carry chains automatically.
- Clock domain
- The set of registers driven by one clock or by clocks with a known, fixed phase relationship. Signals passing between unrelated clock domains need synchronisation.
- Clock domain crossingCDC
- Any signal path from a register in one clock domain to a register in another, unrelated domain. CDC paths need synchronisers, handshakes or asynchronous FIFOs to avoid metastability and data corruption.
- cocotb
- An open-source framework for writing HDL testbenches in Python. Test code runs as coroutines that drive and sample signals in a Verilog or VHDL simulator, so a design can be checked against a Python or NumPy reference model.
- Combinational logic
- Logic whose outputs depend only on its current inputs, with no memory. Multiplexers, decoders, comparators and adders are combinational. In HDL it is written with continuous assignments or always_comb blocks.
- Complex programmable logic deviceCPLD
- A programmable logic device built from a small number of sum-of-products macrocells connected by a central interconnect. CPLDs are usually non-volatile and have predictable pin-to-pin delays, but far less logic than an FPGA.
- Configurable logic blockCLB
- A cluster of LUTs, flip-flops, carry logic and local multiplexers that forms the repeating logic tile of an FPGA. Vendors use different names: CLB and slice (AMD), ALM and LAB (Altera), PFU (Lattice).
- Configuration space
- The space of all joint values of a robot. Planning and control often work in configuration space, while tasks are usually specified in Cartesian task space.
- CORDIC
- COordinate Rotation DIgital Computer: an iterative algorithm that computes rotations, sine and cosine, magnitude and angle using only shifts, additions and a small table of constants, which makes it cheap in FPGA logic.
- Critical path
- The timing path with the least slack. It limits the maximum clock frequency of the design.
- Cyclic redundancy checkCRC
- An error-detecting code computed as the remainder of polynomial division of the message by a generator polynomial. It is cheap in hardware and detects all burst errors shorter than the CRC width.
D
- DAgger
- Dataset Aggregation: an imitation-learning algorithm that runs the learned policy, asks the expert to label the states it visits, and retrains on the growing dataset to reduce compounding errors.
- Degrees of freedomDOF
- The number of independent parameters needed to describe a system's configuration. A rigid body in space has six (three position, three orientation); an arm's DOF usually equals its number of joints.
- Denavit-Hartenberg parametersDH
- A convention that describes each link of a serial arm with four numbers (link length, twist, offset and joint angle), so forward kinematics becomes a product of standard transforms.
- Device under testDUT
- The module being verified by a testbench. Also called the unit under test (UUT).
- Diffusion policy
- A robot policy that generates actions by iteratively denoising random noise, conditioned on observations. It can represent several valid ways of doing a task.
- Direct digital synthesisDDS
- Generating a waveform digitally from a phase accumulator (a numerically controlled oscillator) that indexes a sine look-up table. The frequency is set by the value added to the accumulator each clock.
- Distributed RAM
- Small memories built from the configuration memory of LUTs in logic blocks. Distributed RAM supports asynchronous reads and suits small buffers and register files.
- Domain randomisation
- Varying simulation parameters such as masses, friction, lighting and textures during training, so the policy learns to work across a range that includes the real world.
- DSP slice
- A hard arithmetic block containing a multiplier, pre-adder, accumulator and pipeline registers. DSP slices run multiply-accumulate operations faster and with less power than the same function built from LUTs.
F
- Field-programmable gate arrayFPGA
- An integrated circuit made of configurable logic blocks, memory, arithmetic blocks and programmable routing. Loading a configuration file (bitstream) connects these resources to implement a custom digital circuit, and the device can be reconfigured with a different design later.
- FIFO
- A first-in, first-out buffer with write and read ports and full and empty flags. A synchronous FIFO uses one clock; an asynchronous FIFO uses separate write and read clocks.
- Finite state machineFSM
- A sequential circuit that moves between a finite set of states based on its inputs. In a Moore machine, outputs depend only on the current state; in a Mealy machine, they also depend on the inputs.
- FIR filter
- A finite impulse response filter: each output is a weighted sum of the current and a fixed number of past input samples. On an FPGA each tap maps to a multiply-accumulate, usually in a DSP slice.
- Fixed-point arithmetic
- A number format that stores fractional values as integers with an implied binary point. Q1.15, for example, uses one sign bit and 15 fractional bits. Fixed-point maps directly onto FPGA adders and DSP multipliers.
- Flip-flopFF
- An edge-triggered storage element that captures its D input on a clock edge and holds it until the next edge. In FPGA logic blocks, each LUT output can feed a dedicated D flip-flop with optional clock enable, set and reset.
- Forward kinematicsFK
- Computing the pose of a robot's end effector from its joint values, by chaining the transforms of each joint and link.
G
- Gimbal lock
- The loss of one rotational degree of freedom that happens with Euler-angle representations when two rotation axes line up, so that different angle combinations give the same orientation.
- Gray code
- A binary encoding in which consecutive values differ in exactly one bit. Gray-coded counters can cross clock domains safely because a sample taken during a change is either the old or the new value.
H
- Hardware description languageHDL
- A language for describing digital circuits, such as Verilog, SystemVerilog or VHDL. HDL code describes concurrent hardware and can be simulated or synthesized into gates.
- High-level synthesisHLS
- Tools that generate RTL from an algorithm written in C, C++ or a similar language. Designers guide the result with directives for pipelining, loop unrolling and interfaces.
- Hold time
- The minimum time a flip-flop's data input must remain stable after the active clock edge. Hold violations do not depend on clock frequency and must be fixed by adding delay to the short path.
- Homogeneous transform
- A 4×4 matrix that combines a rotation and a translation, so the pose of one frame relative to another is a single matrix and chains of frames compose by matrix multiplication.
I
- I/O standard
- The electrical specification of an FPGA pin, such as LVCMOS33 or LVDS, covering voltage levels, drive strength and termination. Each I/O bank supports the standards compatible with its supply voltage.
- Imitation learning
- Training a robot to perform a task from demonstrations, for example teleoperated episodes, instead of from a hand-written controller or a reward function.
- Initiation intervalII
- The number of clock cycles between the start of successive loop iterations or function calls in a pipelined design. An II of 1 means a new input every clock.
- Integrated logic analyserILA
- Debug logic inserted into an FPGA design that captures internal signals into on-chip memory when a trigger condition occurs, then reads them out over JTAG. AMD calls it ILA; Altera's equivalent is Signal Tap.
- Intersection over unionIoU
- The area of overlap between a predicted and a true bounding box divided by the area of their union. It is used to decide whether a detection counts as correct.
- Inverse kinematicsIK
- Finding joint values that place the end effector at a desired pose. Simple arms have closed-form solutions; general arms are solved numerically, often with damped least squares.
J
- Jacobian
- The matrix that maps joint velocities to end-effector velocity. It is used for velocity control and numerical inverse kinematics, and it loses rank at singularities.
- JTAG
- A standard serial test and debug interface (IEEE 1149.1). FPGAs use it to load bitstreams, program configuration flash and connect on-chip logic analysers.
K
- Kinematic singularity
- A configuration where the Jacobian loses rank, so the end effector cannot move in some direction and the joint velocities needed near it grow very large.
L
- Latch
- A level-sensitive storage element that is transparent while its enable is active. In FPGA designs, latches are usually created by mistake when a combinational block does not assign an output on every path, and they complicate timing analysis.
- Line buffer
- On-chip memory that stores one or more previous rows of an image so a streaming pipeline can form a 2D window, such as 3×3, around each pixel while processing one pixel per clock.
- Linear-feedback shift registerLFSR
- A shift register whose input is an XOR of selected bits (taps). With the right taps an n-bit LFSR steps through 2^n - 1 states, which makes it useful for pseudo-random sequences, scramblers and counters.
- Look-up tableLUT
- The basic combinational element of most FPGAs. A k-input LUT is a small memory of 2^k configuration bits addressed by its inputs, so it can implement any Boolean function of k variables. Modern devices use 4- to 6-input LUTs.
M
- Markov decision processMDP
- The mathematical model behind reinforcement learning: states, actions, transition probabilities, rewards and a discount factor.
- Maximum clock frequencyFmax
- The highest clock frequency at which a design's critical path meets setup timing, equal to the reciprocal of the minimum achievable clock period.
- Metastability
- A state in which a flip-flop output sits between logic 0 and 1 for an unpredictable time after its setup or hold window was violated. The probability that it persists falls exponentially with the time allowed to resolve.
O
- Object detection
- Locating and classifying objects in an image, usually as bounding boxes with class labels and confidence scores.
- One-hot encoding
- A state encoding that uses one flip-flop per state, with exactly one bit set at a time. It uses more flip-flops than binary encoding but often gives simpler, faster next-state logic in FPGAs.
P
- Phase-locked loopPLL
- A clock management circuit that generates output clocks at a multiple or fraction of an input clock, with controlled phase. FPGAs include PLLs or similar blocks (AMD calls one variant an MMCM) for clock synthesis and deskew.
- Physical unclonable functionPUF
- A circuit whose output depends on uncontrollable manufacturing variation, so each chip responds to a challenge differently. PUFs are used to derive device-unique keys or identifiers.
- PID controller
- A feedback controller whose output is a weighted sum of the error (P), its integral (I) and its derivative (D). It is the standard loop for joint position and speed control.
- Pipelining
- Inserting registers into a long combinational path to split it into shorter stages. Pipelining raises the achievable clock frequency and throughput at the cost of added latency.
- Place and routeP&R
- The implementation step that assigns each netlist primitive to a physical location on the FPGA (placement) and connects them through the programmable routing (routing), guided by timing constraints.
Q
- Quadrature phase-shift keyingQPSK
- A digital modulation that sends two bits per symbol by choosing one of four carrier phases, using an in-phase (I) and a quadrature (Q) component. BPSK is the one-bit, two-phase version.
- Quantisation
- Representing a neural network's weights and activations with fewer bits, for example 8-bit integers, to cut memory and speed up inference on edge hardware, at a small cost in accuracy.
R
- Register
- A group of flip-flops that share a clock and store a multi-bit value. In HDL, a register is inferred from a signal assigned inside a clocked process, such as an always_ff block.
- Register-transfer levelRTL
- A level of design abstraction that describes a circuit as registers and the combinational logic that moves and transforms data between them on each clock cycle. Most FPGA design is written at RTL.
- Reinforcement learningRL
- Learning a policy by trial and error to maximise expected cumulative reward. Robot RL is mostly done in simulation because real trials are slow and can damage hardware.
- Rotation matrix
- A 3×3 orthonormal matrix with determinant +1 that rotates vectors and describes the orientation of one coordinate frame relative to another. Its columns are the rotated frame's axes expressed in the reference frame.
S
- Servo motor
- A motor with built-in position sensing and a control loop that drives it to a commanded position. Smart bus servos also report position, speed and load over a shared serial bus.
- Setup time
- The minimum time a flip-flop's data input must be stable before the active clock edge for the value to be captured reliably. Setup violations are fixed by shortening the path or lowering the clock frequency.
- Sim-to-real transfer
- Training or testing a robot policy in simulation and then running it on real hardware. Domain randomisation and system identification help close the gap between the two.
- Slack
- The margin by which a timing path meets its requirement: required time minus arrival time. Positive slack means the path passes; negative slack means it fails. Reports summarise this as WNS (worst negative slack) and TNS (total negative slack).
- Soft processor
- A CPU core built from FPGA logic rather than fixed silicon, so its options can be configured per design. A hard processor is a fixed block on the same chip.
- Static timing analysisSTA
- A method that checks every timing path in a design against its constraints without simulation. STA reports setup and hold slack for each path and flags paths that violate their requirements.
- Synchronizer
- A chain of two or more flip-flops in the destination clock domain that gives a metastable value time to resolve before the rest of the logic uses it. It is only safe for single-bit signals or Gray-coded values.
- Synopsys Design ConstraintsSDC
- A Tcl-based format for timing constraints such as create_clock, set_input_delay and set_false_path. Quartus uses SDC directly, and AMD's XDC format is based on it.
- Synthesis
- The step that converts RTL code into a netlist of device primitives such as LUTs, flip-flops, block RAM and DSP slices. Synthesis also optimises logic and reports inferred resources and warnings.
- SystemVerilogSV
- The successor to Verilog, standardised as IEEE 1800. It adds design features such as logic, always_ff, always_comb, enums and interfaces, and a large set of verification features including classes, constrained randomisation and assertions.
- SystemVerilog assertionsSVA
- A part of SystemVerilog for stating properties a design must satisfy, either as immediate checks or as temporal properties over clock cycles. Assertions are checked in simulation and by formal verification tools.
- Systolic array
- A grid of identical processing elements that pass data to their neighbours on every clock, so each value fetched from memory is reused many times. Matrix multiplication for neural networks is a common use.
T
- Teleoperation
- Controlling a robot remotely in real time. In leader-follower setups a person moves a leader arm and the follower arm copies its joint positions, which is how many robot-learning datasets are recorded.
- Testbench
- HDL code that drives a design under test with stimulus in simulation and checks its outputs. A self-checking testbench compares results against expected values and reports pass or fail without manual waveform inspection.
- Timing closure
- The process of modifying a design, its constraints or the tool settings until all timing paths meet their setup and hold requirements.
- Transition-minimised differential signallingTMDS
- The serial link used by DVI and HDMI. Each 8-bit colour value is encoded into a 10-bit symbol that reduces transitions and keeps the line DC-balanced, then sent on a differential pair.
- True random number generatorTRNG
- A generator that draws randomness from a physical noise source, such as jitter in ring oscillators, rather than from a deterministic algorithm. Its output must be monitored by health tests.
U
- UART
- Universal asynchronous receiver-transmitter: a serial interface that sends bytes framed by a start bit and stop bits, with no shared clock. Both ends agree on a baud rate, such as 115200 bits per second.
- Unit quaternion
- A four-number representation of 3D orientation, q = (w, x, y, z) with unit length. Quaternions avoid gimbal lock, compose by multiplication and interpolate smoothly with SLERP.
V
- Verilog
- A hardware description language with a C-like syntax, standardised as IEEE 1364 and later merged into SystemVerilog (IEEE 1800).
- VHDL
- A strongly typed hardware description language standardised as IEEE 1076. It is common in European industry, aerospace and defence. The name stands for VHSIC Hardware Description Language.
- Video timing
- The sequence of active pixels, front porch, sync pulse and back porch on every line and frame of a raster display, all counted in pixel clocks. A video controller generates it with horizontal and vertical counters.
- Vision-language-action modelVLA
- A neural network that takes camera images, a natural-language instruction and the robot's state, and outputs robot actions. VLAs are usually built on pretrained vision-language models.
X
- Xilinx Design ConstraintsXDC
- The constraint file format used by AMD Vivado. It combines SDC-style timing constraints with physical constraints such as pin locations and I/O standards.