Glossary

Action chunk. A short sequence of future actions predicted at once and executed before the next prediction.

Action expert. A smaller network that generates continuous actions (often with flow matching) while attending to a VLM's features or key-value cache, so the large model runs once per chunk rather than once per action token.

ADE / minADE. Average displacement error: mean distance between a predicted and a logged trajectory. minADE over k samples takes the best of k predictions.

Advantage. How much better an action did than expected: return (or reward) minus a baseline such as the state's value.

AlpaSim. The closed-loop driving simulator the Alpamayo-R1 paper evaluates in.

Asynchronous inference. Starting the next model inference while the current action chunk is still being executed, so the robot does not wait. The late-arriving plan must have its elapsed waypoints dropped.

Audit sample size. How many labels a human must check to estimate an error rate within a margin: n = 1.96² p(1−p) / margin². It does not depend on the dataset size, only on the margin and the error rate, and the sample must be random.

Bellman equation. A state's value equals the expected reward now plus the discounted value of the next state; the basis of value iteration and Q-learning.

BEV (bird's-eye view). A top-down grid of features around the vehicle, used as a shared space for fusing sensors.

Bicycle model. A vehicle model with one front and one rear wheel: heading changes at speed times tan(steering angle) over wheelbase. Used to check that a trajectory is drivable.

Bootstrap interval. An uncertainty range from resampling the labelled items with replacement and recomputing the statistic (for example kappa) many times.

Bradley–Terry model. P(response A preferred to B) = sigmoid(r(A) − r(B)); the standard way to train a reward model from comparisons.

Catastrophic forgetting. A network trained on a new task loses much of what it learned for an earlier one, because the same weights are reshaped. Limited by replay, penalties on important weights (EWC), small steps and adapters.

Chain of Causation (CoC). A label style that ties a driving decision to its observed cause.

Chain of thought (CoT). Having a model write intermediate reasoning steps before its answer.

Classifier-free guidance. Steering a conditional diffusion model by combining its conditional and unconditional noise estimates.

Close-encounter rate. How often a planned path comes within an unsafe distance of another agent.

Closed loop / open loop. Closed loop means the model's own actions change what it sees next. Open loop compares predictions to a recording without acting on them.

Cohen's kappa. Agreement between two labellers beyond what chance would give: (observed − chance) / (1 − chance). A dominant class can make raw agreement high and kappa low.

Common cause failure. A single cause that disables redundant channels together (shared power, shared software, the same weather). It limits the benefit of redundancy: with a common share β, both channels fail with probability about βp instead of p².

Conformal prediction. A method that turns any predictor's calibration errors into intervals with a coverage guarantee, assuming exchangeable data.

Data engine. A loop that runs the current model, finds where it fails, labels those cases, adds them to training and repeats. Its value comes from targeting failures instead of sampling data at random.

Data parallelism (DDP). Every GPU holds a full model copy and a slice of the batch; gradients are averaged across GPUs after each backward pass.

DDIM. A deterministic diffusion sampler that reuses a trained DDPM network with far fewer steps.

Discount factor (gamma). A number between 0 and 1 that weights later rewards less in the return.

Distillation. Training a small or fast student to imitate a large or slow teacher's outputs, features or actions.

DPO. Direct Preference Optimization: training a policy directly on preferred versus rejected outputs, without a separate reward model.

End-to-end. One learned model from sensors to trajectory, instead of separate perception, prediction and planning modules.

Epsilon-greedy. Act on the best-looking choice, except with probability epsilon pick at random, to keep exploring.

Extrinsics / intrinsics. Extrinsics: where a camera is mounted and how it points relative to the vehicle. Intrinsics: its focal lengths and principal point.

FAST. Frequency-space Action Sequence Tokenization: compresses an action chunk with a discrete cosine transform before tokenising it.

Flow matching. A way to generate continuous outputs by learning a velocity field that moves noise toward the data distribution. Related to diffusion.

FSDP / ZeRO. Sharded data parallelism: optimiser state, gradients and (at stage 3) weights are split across GPUs instead of copied.

FSQ. Finite scalar quantization: rounding each latent dimension to a few fixed levels to make discrete tokens without a learned codebook.

GRPO. Group Relative Policy Optimization: a reinforcement learning method that scores several sampled outputs for the same input against each other.

Hungarian matching. The lowest-cost one-to-one assignment of predicted objects to ground-truth objects, used so a set loss does not depend on output order.

Hybrid stack. A driving system that keeps some modular components (often perception) and learns the rest, or uses module outputs as auxiliary tasks.

IDM (Intelligent Driver Model). A rule-based car-following formula that accelerates toward a desired speed and brakes when the gap to the leader is below the gap it wants.

Innovation. In a Kalman filter, the measurement minus what the filter predicted for it. Its normalised square (NIS) averages about 1 when the noise settings are honest.

Joint prediction. Predicting the futures of several agents together, so sampled futures are consistent with each other. A marginal predictor samples each agent separately and can produce futures in which they collide.

Kalman filter. A recursive estimator that alternates predicting the state with a motion model and correcting it with a measurement, weighing the two by their uncertainties.

Keyframe. The moment in a clip chosen for labelling, typically just before the behaviour changes, so the labeller sees the cause without the result.

KV cache. Stored keys and values of earlier tokens, reused so each decoding step does not recompute them.

Latency budget. The deadline for one inference, split across the vision encoder, prefill, reasoning decoding and action decoding.

Lift-splat. Turning image features into a bird's-eye-view map by predicting a depth distribution per pixel, placing the feature at each candidate depth weighted by its probability, and summing into grid cells.

Long tail. Rare situations that are individually uncommon but collectively frequent.

LoRA. Low-rank adaptation: freezing a pretrained weight and training a small low-rank update added to it.

MDP (Markov decision process). States, actions, rewards and transition probabilities that depend only on the current state and action.

Memory-bound. Limited by the time to read data from memory rather than by arithmetic; single-sequence token decoding is, because every weight is read once per token.

Meta-action. A coarse behaviour label such as accelerate, cruise or decelerate, derived from the ego log by rules.

MFU (model FLOPs utilisation). The share of the GPUs' peak arithmetic throughput a training run actually achieves.

NAVSIM. A benchmark that evaluates driving policies on real data with a short non-reactive simulation, between open-loop and closed-loop testing.

Object query. A learned vector that attends to the input and produces one predicted object; a fixed set of them predicts a whole scene (DETR).

Offline RL. Reinforcement learning from a fixed dataset, with no further interaction.

PDM score (PDMS). NAVSIM's combined driving score: penalty terms for collisions and leaving the road multiply a weighted average of progress, time to collision and comfort.

Pillar. A vertical column of space over one BEV cell; lidar encoders such as PointPillars pool the points in each pillar into one feature vector.

Polyline encoder. A map or agent-history encoder that turns each polyline into relative vectors, embeds them and max-pools, so the result does not depend on where the polyline sits in the world.

PPO. Proximal Policy Optimization: a policy-gradient method that clips the change in action probabilities and uses a learned value network as the baseline.

Prefill. The first pass of a language model over the whole prompt (image tokens, route, text) in parallel, which fills the KV cache before decoding starts.

Process supervision. Training signal that scores each reasoning step rather than only the final answer.

Q-learning. Learning action values Q(s, a) from experience by moving each estimate toward reward plus discounted best next value.

Rarity score. For mining the long tail: the mean distance from a scene to its k nearest neighbours in an embedding space. k must exceed the size of the rare clusters to be found.

Reality gap. The mismatch between a simulator and the real world that makes simulated policies fail on hardware.

Replay. Mixing examples of an old task into the training for a new one so the weights keep serving both. Called co-training when the old data is a standing part of the recipe.

Return. The sum of (discounted) rewards from a point in an episode onward; what an RL agent maximises.

Reward hacking. A policy scoring well on the written reward without achieving the intended goal.

Reward model. A network trained on human (or AI) comparisons to score outputs; the reward in RLHF.

RLAIF. RL from AI feedback: the comparisons that train the reward model come from a model instead of people.

RLHF. RL from human feedback: SFT, a reward model from human comparisons, then RL against it with a KL penalty.

RLVR. RL with verifiable rewards: the reward comes from a checker (a correct answer, passing tests), not a learned model.

Route conditioning. Giving a driving policy the route (a command, waypoints, a goal or text) so it can choose a branch that the camera image does not decide.

RSS (Responsibility-Sensitive Safety). A formal model of careful driving. Its longitudinal rule gives the smallest gap from which a following car can always stop under worst-case assumptions about the car ahead.

Rule of three. After n failure-free trials, the 95 percent upper bound on the failure rate is about 3/n; so showing a rate below r needs about 3/r failure-free miles.

Scaling law. A power-law relation between loss and data, model size or compute, used to predict larger runs from smaller ones.

Self-consistency. Sampling several reasoning paths and choosing the most common answer.

SFT. Supervised fine-tuning: training on labelled examples with the usual loss.

Token pruning. Removing image tokens at inference, before or inside the language model, to cut prefill and KV-cache cost.

Trajectory vocabulary. A large fixed set of plausible trajectories, usually clustered from logs, that a planner scores or chooses among.

Unicycle model. A vehicle model controlled by acceleration and steering curvature (or equivalent), used so predicted trajectories are physically followable.

Value function. Expected return from a state (V) or from a state and action (Q) under a policy.

Vector quantization (VQ). Replacing a latent vector with its nearest entry in a learned codebook; the entry's index is the token.

VLA. Vision-language-action model: a pretrained vision-language model extended to output actions.

VLM. Vision-language model: takes images and text, produces text.

World action model (WAM). A model that forecasts future observations and outputs actions; recent ones use the forecast only as a training signal.

World model. A model that predicts how the world changes after an action, used to plan, to train on imagined experience, or to simulate.