Learning Human-Like Car-Following Behaviour for ACC

The problem

Drivers often find ACC unnatural: it may brake too early, accelerate too slowly or keep gaps that invite cut-ins, so many switch it off. Human drivers adapt their following style to traffic, speed and road type. Learning this behaviour from data could make ACC more acceptable, but learned models can also learn unsafe habits and behave unpredictably outside their training data. Combining learned style with guaranteed safety is a research problem with practical relevance.

Objectives

• Extract car-following episodes from naturalistic data and from driving-simulator sessions.
• Calibrate classical car-following models such as IDM and Gipps as baselines.
• Train learned car-following models, for example a neural network or Gaussian process.
• Wrap the learned model in a safety layer that enforces minimum gap and deceleration limits.
• Evaluate imitation accuracy, safety, comfort and behaviour outside the training distribution.

What you need

Key concepts

Car-following model

A model that gives the follower's acceleration from its speed, the gap and the relative speed to the leader. IDM and Gipps are classical examples with a few interpretable parameters.

Behaviour cloning

Training a model to reproduce recorded human actions from the same inputs. It is simple but can drift into unseen states when run in closed loop.

Closed-loop evaluation

Testing a learned model by letting it drive in simulation, not just predicting the next acceleration. Small prediction errors can accumulate into unrealistic or unsafe behaviour.

Safety layer

A filter that checks the learned acceleration against safety constraints, such as a minimum time gap and maximum deceleration, and overrides it when needed.

Step-by-step guide

1. 1

   Frame questions and ethics

   Write the research questions: can a learned model imitate human car-following more accurately than calibrated IDM and Gipps models, and does a safety layer keep it safe without destroying the human-like feel? Plan the simulator study with informed consent, anonymised data and approval from your institute's ethics committee if required. State that no learned controller will be tested on public roads, only in simulation.

   Checkpoint: A research plan with questions, metrics, data sources and the ethics arrangements for simulator sessions.

2. 2

   Extract naturalistic car-following episodes

   From the Waymo Open Dataset motion data, identify episodes where one vehicle follows another in the same lane for at least 10 to 15 seconds. Compute gap, speeds, relative speed and follower acceleration at a fixed rate, smoothing positions before differentiating. Filter out lane changes and implausible values. Split by scenario, not by time step, into training, validation and test sets, so the same episode never appears in two sets.

   Checkpoint: A cleaned episode dataset with summary statistics and a scenario-level split.

3. 3

   Collect simulator data

   In CARLA, build following scenarios with a scripted lead vehicle: steady following, stop-and-go, lead braking and cut-ins, on urban and highway maps. Recruit volunteer drivers to drive with a steering wheel and pedals after a practice session. Log gap, speeds and accelerations at a fixed rate. Keep each driver's data separate so you can study individual styles and test generalisation to new drivers.

   Checkpoint: Simulator logs from several drivers with per-driver summaries of gap and acceleration.

4. 4

   Calibrate classical baselines

   Calibrate IDM and Gipps parameters for each dataset by minimising the error between simulated and recorded gap over whole episodes, not one-step acceleration error, using an optimiser such as SciPy's differential evolution. Calibrate both a global set and per-driver sets for the simulator data. Report calibrated parameters with their ranges and check they are physically plausible. Repeat calibration from several starting points to check the optimum is stable.

   Checkpoint: Calibrated IDM and Gipps parameters with closed-loop gap error on validation episodes.

5. 5

   Train learned models

   Train a feed-forward network on current gap, speed and relative speed, and a recurrent network (LSTM or GRU) on a short history of these, both predicting follower acceleration. Use the validation set for early stopping and hyperparameter choice. Optionally fit a Gaussian process for uncertainty estimates. Record training settings and seeds so every model can be reproduced exactly. Normalise inputs with training-set statistics only, and keep the test set untouched until the final evaluation.

   Checkpoint: Trained models with one-step prediction error on validation data compared with the baselines.

6. 6

   Evaluate in closed loop

   For every test episode, replay the recorded leader trajectory and let each model drive the follower in simulation from the recorded initial state. Compare simulated gap and speed with the recorded human trajectory using RMSE of gap and speed. Also count collisions, gaps below a safety threshold and unrealistic accelerations. Closed-loop testing often reverses one-step rankings, so report both. Report results for naturalistic and simulator data separately.

   Checkpoint: Closed-loop gap and speed RMSE and safety counts for all models on the test set.

7. 7

   Design the safety layer

   Wrap the learned model's acceleration with a safety filter: compute the acceleration bound that keeps a minimum time gap under a worst-case lead deceleration assumption, and clip the learned output to it and to comfort limits. Optionally implement it as a simple quadratic programme that stays as close as possible to the learned output while meeting constraints. Log every intervention.

   Checkpoint: The filtered learned controller shows zero collisions in closed-loop tests, with interventions logged.

8. 8

   Integrate into an ACC simulation

   Export the learned model to ONNX and call it from Simulink, or reimplement the network in a MATLAB Function block. Combine it with the safety layer and Stateflow mode logic for speed and following modes. Run the same scenario library as for a conventional constant time-gap ACC, including cut-ins and hard braking, at several speeds. Check that the Simulink outputs match Python outputs on the same inputs before running the library, so any difference later comes from the closed loop and not from export errors.

   Checkpoint: The learned ACC runs in Simulink on the scenario library with the safety layer active.

9. 9

   Compare behaviour and comfort

   Compare learned ACC, IDM-based ACC and conventional ACC on human-likeness (closed-loop error against held-out human trajectories), safety (minimum gap and TTC), comfort (RMS jerk) and safety-layer intervention rate. Optionally ask simulator volunteers to ride along in a passive mode and rate naturalness, recording ratings without implying any statistical claim beyond your sample. Use the same scenarios, seeds and safety limits for every controller in the comparison.

   Checkpoint: A comparison table across controllers with human-likeness, safety, comfort and intervention rate.

10. 10

    Test outside the training data

    Test the models on situations missing or rare in training: higher speeds, very short cut-ins, a stationary vehicle ahead and a lead vehicle reversing slightly. Record how often the safety layer must intervene and whether the learned model behaves erratically. This shows the risk of learned controllers and justifies the safety layer. Compare with IDM, whose simple structure usually extrapolates more predictably, and describe each erratic case with plots of gap and acceleration.

    Checkpoint: Out-of-distribution results with intervention rates and examples of erratic behaviour.

11. 11

    Write up the research

    Present the data extraction, simulator study, baselines, learned models, closed-loop evaluation, safety layer, ACC integration and out-of-distribution tests. Discuss whether human-like behaviour is desirable when humans also follow too closely, and where the safety layer limits imitation. Release code and processed episode lists where dataset terms allow. State clearly which results come from naturalistic data and which from the simulator, since simulator driving can differ from real driving. Include the consent and anonymisation procedure in an appendix.

    Checkpoint: Thesis chapters with closed-loop results, safety analysis and released code.

Testing & validation

• Calibrated IDM and Gipps baselines with closed-loop error on validation episodes.
• Closed-loop gap and speed RMSE for learned models on held-out data.
• Zero collisions with the safety layer, with intervention rate reported.
• Comfort and safety metrics compared with conventional ACC.
• Behaviour and intervention rate in out-of-distribution scenarios.

Expected results

A learned, human-like car-following ACC with a safety layer, evaluated in closed loop against classical models and conventional ACC, with a clear account of its imitation accuracy, safety and limits.

Take it further

• Personalise the ACC to an individual driver's style with a small amount of their data.
• Use inverse reinforcement learning to learn a reward instead of cloning actions.
• Add lane-change and cut-in prediction from neighbouring vehicles.
• Study driver acceptance with a larger simulator panel.

Viva and review questions

Learn it at DIYguru

These DIYguru courses and programmes teach what this project needs.