Strategy 3: Real-Time Correction¶
In the V3 review
Literature scenarios that sense and correct mid-process (Fig. 3):
- watch → detect → correct: computer vision corrects over- and under-extrusion in bioprinting within ~10 s [33].
- scoop → verify → pour: vision-guided deep-RL powder scooping with real-time volume correction, ~1.93 mg error across seven powders [12].
See Literature Scenarios.
Overview¶
Sense the process as it happens, adjust parameters mid-execution to keep outcome on target.
Real-time correction monitors a fast process (extrusion, compression, filling) and adjusts the control input (pressure, force, speed) continuously to maintain the desired trajectory.
When to Use¶
- Process is unpredictable (yield-stress material, variable cohesion, unclear flow initiation)
- Fast feedback available (pressure transducer, load cell, camera at 30+ fps)
- Errors compound quickly (if you miss target early, can't recover by end)
- Material changes mid-process (e.g., powder bridge breaks partway through filling)
- Safety-critical or value-critical (soft tissues, expensive reagents)
- Time budget is moderate (process takes 5–30 seconds)
Common applications: - Bio-printing (soft inks that yield and strain) - Fragile object handling (tissue, foam, delicate electronics) - High-accuracy powder dosing with live weight feedback - Robot compliance control (feel-based grip adjustment)
How It Works¶
Step 1: Define the Desired Trajectory¶
Specify what you want to measure over time. Examples:
- Fill height (camera): target trajectory is linear: h(t) = h_target × (t / t_total)
- Weight (balance): target is linear: W(t) = W_target × (t / t_total)
- Pressure (transducer): target is constant: P(t) = P_setpoint
- Force (load cell): target is constant, or follows a ramp: F(t) = F_target
Step 2: Set Up Fast Feedback Loop¶
Sensor bandwidth must be at least 10× faster than process dynamics.
- Extrusion process: 5–30 second duration → sensor sampling ≥ 100 Hz
- Grip and retrieval: 0.5 second duration → sensor sampling ≥ 200 Hz
- Soft material compression: 1–5 seconds → sensor sampling ≥ 100 Hz
Step 3: Implement Proportional or PID Control¶
At each sensor sample time (every 10–100 ms), compute:
Adjust the control parameter (pump speed, pressure, force) proportionally:
Where \(K_p\) (proportional gain) is tuned to avoid oscillation and overshoot.
PID (More Sophisticated)¶
For better stability, use Proportional-Integral-Derivative:
- \(K_p\) (proportional): Respond to current error
- \(K_i\) (integral): Correct steady-state drift
- \(K_d\) (derivative): Dampen overshoot
Step 4: Monitor for Completion¶
Stop dispensing/handling when: - Measured outcome ≈ target - Time limit exceeded (safety failsafe) - Sensor reading saturates or fails (fault detection)
Advantages¶
| Aspect | Benefit |
|---|---|
| Robustness | Adapts to surprises mid-action; compensates for unmeasured drift |
| Speed | No iterative cycles; completes in one pass (5–30 seconds) |
| Handling difficult materials | Works with yield-stress, non-Newtonian, unpredictable materials |
| Safety | Can detect overload or unexpected force and stop immediately |
| Accuracy | Can achieve ±2–5% accuracy in real time |
Limitations¶
| Challenge | Issue |
|---|---|
| Sensor latency | If sensor delay > 500 ms, feedback control becomes unstable (oscillates) |
| Tuning complexity | PID gain tuning is an art; too high causes ringing, too low is sluggish |
| Computational overhead | Requires fast loop rate (100+ Hz) and fast computer |
| Sensor noise | Noisy measurement causes noisy corrections; needs filtering |
| Non-linear systems | Linear PID may not work for highly non-linear processes |
Real-World Examples¶
Example 1: Powder Filling with Live Weight Feedback¶
Task: Fill a small vial with exactly 2.0 g of powder
Setup: - Auger feeder on robot arm - Vial sits on a precision load cell (0.1 g resolution) - Dispense at speed 20 rpm (nominal) - Control loop runs at 10 Hz (read weight every 100 ms)
Execution:
Time (s) | Weight (g) | Error (g) | Speed (rpm) | Action
0.0 | 0.000 | +2.000 | 20 | start
0.5 | 0.200 | +1.800 | 20 | continue
1.0 | 0.500 | +1.500 | 20 | continue
2.0 | 1.200 | +0.800 | 20 | continue
3.0 | 1.850 | +0.150 | 20 | continue
3.5 | 1.980 | +0.020 | 18 | slow down
4.0 | 2.005 | −0.005 | 15 | nearly done
4.1 | 2.008 | −0.008 | stop | STOP (within tolerance)
Proportional gain: K_p = 5 rpm/g (for every 0.1 g excess, reduce speed by 0.5 rpm)
Outcome: Filled in 4.1 seconds, final weight 2.008 ± 0.005 g. ✓
Example 2: Bio-Printing with Pressure Control¶
Task: Extrude 10 µL of a delicate cell suspension from a syringe.
Setup: - Air-pressure syringe: nominal 80 kPa - Pressure transducer on syringe barrel (±1 kPa accuracy) - Extrusion nozzle sits 5 mm above substrate - Camera monitors extrusion rate in real time
Execution:
The material is shear-sensitive. If pressure is too high, cells are damaged. If too low, flow stops.
Time (s) | Pressure (kPa) | Flow Rate (µL/s) | Target Flow | Action
0.0 | 80 | 0 (blocked) | 10/5=2 | pressure too high for this batch!
0.5 | 70 | 1.2 | 2.0 | below target
1.0 | 60 | 0.9 | 2.0 | still below
1.5 | 75 | 1.8 | 2.0 | close
2.0 | 78 | 2.1 | 2.0 | close
3.0 | 76 | 2.0 | 2.0 | ✓ stable
5.0 | 76 | 2.0 (5 µL) | 10 µL total | half done
...
Control law: If flow_rate > target, reduce pressure by 2 kPa; if below, increase by 1 kPa
Outcome: Extrusion completed at 5 seconds, 10 ± 0.5 µL (cells survive) ✓
Example 3: Robotic Grip Adjustment for Delicate Solid¶
Task: Grip and lift a soft foam block without crushing it
Setup: - 6-DOF robot with 2-finger gripper - Force/torque sensor in wrist - Gripper finger displacement measured via encoder
Execution:
Time (ms) | Grip Force (N) | Displacement (mm) | Target Force | Action
0 | 0 | 0 | 5 N | start closing
200 | 1 | 2 | 5 N | continue
400 | 3 | 4 | 5 N | continue
600 | 5 | 5.5 | 5 N | ✓ on target
800 | 5.2 | 5.6 | 5 N | force stabilized
1000 | 5 | 5.7 | 5 N | ✓ hold and lift
2000 | 5 | 5.7 | 5 N | moving to drop location
Control: For every 0.5 N excess force, open gripper 0.1 mm
Outcome: Foam gripped and lifted without crush damage ✓
Practical Implementation: PID Tuning¶
Gain Selection¶
- Start with P only (I = D = 0):
- Set K_p = 1 (conservative)
- Run a trial
- If response is slow, increase K_p to 2, 5, 10, etc.
- If response oscillates, decrease K_p
- Add I term if there's steady-state error:
- Set K_i = K_p / 10 (rule of thumb)
- Reduces long-term drift
- Add D term if oscillation persists:
- Set K_d = K_p / 100
- Dampens overshoot
Ziegler-Nichols Method (Classic)¶
- Set K_i = K_d = 0
- Increase K_p until system oscillates at steady amplitude and frequency
- Note the critical K_p value and oscillation period T_osc
- Set:
- K_p_final = 0.6 × K_p_critical
- K_i = K_p_final / T_osc
- K_d = K_p_final × T_osc / 8
Practical Tuning Tips¶
- Noisy sensor? Add a low-pass filter (e.g., exponential moving average)
- Oscillating? Reduce K_p or increase K_d
- Slow to reach target? Increase K_p (carefully)
- Overshoots? Reduce K_p or increase K_d
Implementation Checklist¶
- Identify the quantity to control (weight, height, force, pressure)
- Select a sensor with adequate accuracy and speed
- Verify sensor latency is < 100 ms for a 5–30 second process
- Design a target trajectory (constant, linear, or time-varying)
- Implement PID loop in control software
- Perform initial tuning on 3 test materials
- Set safety limits (max/min control signal, timeout)
- Test on production material
- Document K_p, K_i, K_d gains for reference
- Monitor system performance over time (drift, wear)
Comparison with Other Strategies¶
| Strategy | Reaction Time | Accuracy | Material Robustness | Complexity |
|---|---|---|---|---|
| Characterize, then act | Slow | ±5% | Low | Low |
| Iterative optimization | Slow (offline) | ±1% | Medium | Medium |
| Real-time correction | Very fast | ±2% | Very high | High |
| Feedforward prediction | Fast | ±2% | Medium | High |
| Simulation generalization | Fast | ±5% | High | Very high |
References¶
- Rastegardoost, M. M.; Bao, Z.; Sadr, S.; Abdollahi, N.; Darvish, K.; Clasky, A.; Ohsaka, H.; Le Devedec, F. Self-Driving Laboratories for Formulation: Handling Liquids, Pastes, Powders and Objects, from Robots to Digital Twins and AI Agents. Manuscript in preparation, 2026.
- Key references from the review: [12], [33]. Full list on the References page.
Further reading (from the earlier wiki)¶
- Åström, K. J., & Murray, R. M. (2010). Feedback Systems: An Introduction for Scientists and Engineers. Princeton University Press.
- Ogata, K. (2010). Modern Control Engineering (5th ed.). Prentice Hall.