E1 · Measured latency
3 minApply the step to the knee in both worlds.
Done when: python3 kit/atraso.py --medir → atraso medido (real − sim): 22 ms.
Lesson 4 · 16 min
The real actuator: latency, torque limit and actuator network
Between the policy deciding and the motor turning there is a network, the controller loop and the motor PD. In this lesson you measure the "real robot" latency (delay) with a step on the knee, count the falls that 20 and 40 ms cause under a push, and compensate the kp identified in lesson 3.
Start the lesson
In this lesson you will
- 1The layers of time: policy at 50 Hz, network, PD at 500 Hz–1 kHz
- 2Native latency in MuJoCo 3.5+: the actuator's delay attribute
- 3Falls under a push with 0, 20, and 40 ms of latency
- 4Compensate the identified kp and watch the "real" robot walk
Goal: measure the actuator latency on the bench, count the falls it causes under a push, and compensate kp in the command.
01The path from command to torque
Click each block.
Click each block in the diagram to see what it does.
Latency and saturation are invisible in a calm walk and show up under a push.
02Three ways to model the actuator
From the simplest to what you use when the simple one is not enough.
1. Ideal PD
What MuJoCo and Isaac Lab do without you asking: the torque comes out instantly, exactly equal to the PD. It is the "sim" from lesson 1.
2. PD + latency + limit
Adds the measured latency and the declared torque limit. MuJoCo 3.5+ handles the latency by itself (delay and nsample on the actuator); the kit fills the history with the first command so it does not start from a zero target.
With sysid (lesson 3) and randomization around the measured value, this is what many projects do.
3. Actuator network
A small neural network learns, from real motor data, the relationship between command, position history, and delivered torque, and enters the simulator in place of the ideal PD.
Worth it when the actuator has effects that a single parameter does not capture. It requires a motor test bench; this course only mentions it.
4. Gain per motor
NVIDIA's GR00T-WholeBodyControl stack added, on 08/31/2026, per-motor kp and kd scaling against stumbling in the SONIC policy: sim-to-real is also fine-tuning the PD at deploy time.
E3 in this lesson does the minimal version: it multiplies the commanded kp by the inverse of the identified one.
03Measure the latency and count the falls
--medir gives a 0.3 rad step at the knee with the robot hanging. --ms runs the walk with 6 pushes (120, 150, and 180 N for 0.2 s, forward and sideways) per latency.
Terminal · measure the latency on the bench
F=${FTH:-$HOME/fth}; cd $F/sim2real && . $F/.venv/bin/activate
python3 kit/atraso.py --medirTerminal · falls by latency
F=${FTH:-$HOME/fth}; cd $F/sim2real && . $F/.venv/bin/activate
python3 kit/atraso.py --ms 0 20 40Example from the repository
degrau de 0,3 rad no joelho esquerdo, robô pendurado: tempo até o encoder mexer 0,005 rad sim: 8 ms real: 30 ms atraso medido (real − sim): 22 ms
The "real" takes 22 ms longer than the sim to move the knee: the 20 ms of hidden latency plus a bit from the softer kp, which also slows the response.
atraso quedas em 6 empurrões (120/150/180 N, frente e lado, 0,2 s, t = 3 s) 0 ms 2/6 . . . . x x 20 ms 3/6 x x x . . . 40 ms 6/6 x x x x x x (x = caiu, . = ficou de pé; ordem: 120, 150, 180 N para a frente e depois para o lado)
At 40 ms, it falls in all six. At 0 ms, it still falls in two side pushes; at 20 ms, in the three front ones.
Don't conclude that "20 ms is better from the side than 0 ms." These are six trials with one seed: the fall pattern wobbles; what you read is the count rising with latency (2, 3, 6). To claim more, add more pushes and directions.
04Compensate kp in the command
Lesson 3 identified kp ≈ 0.57 of the commanded value. Sending 1/0.57 ≈ 1.75 times the kp, the "real" motor delivers close to what the policy expects.
Terminal · compensated kp
F=${FTH:-$HOME/fth}; cd $F/sim2real && . $F/.venv/bin/activate
python3 kit/dois_mundos.py --kp-comando 1.75; echo "saída: $?"Example from the repository
política: unitree_rl_lab G1 29 GDL velocity v0 (ONNX, 480 entradas) · comando 0.50 m/s · 10 s sim queda: não · andou 5.1 m em 10 s (vel. 0.51 m/s; pedida 0.50) real simulado queda: não · andou 4.5 m em 10 s (vel. 0.45 m/s; pedida 0.50) saída: 0
The "simulated real" now walks 4.5 m and doesn't fall (exit 0). The sim, with the stiffer kp, keeps walking. The latency, the mass, and the friction are still there: the policy handles it, because the biggest difference was corrected.
05Practice
Actuator timing and torque.
E2 · Falls from latency
5 minRun the walk with pushes at 0, 20, and 40 ms.
Done when: python3 kit/atraso.py --ms 0 20 40 → 0 ms 2/6, 20 ms 3/6, 40 ms 6/6.
E3 · Compensate and walk
5 minUse the kp you identified in lesson 3 (1 divided by it) and run both worlds again.
Done when: python3 kit/dois_mundos.py --kp-comando 1.75; echo "saída: $?" → real simulado … queda: não and saída: 0.
06Lesson quiz
The explanation appears after your answer.
Question 1 of 3
With no push, the policy walks with the "real" robot's 20 ms of latency (lesson 1). With a push, 40 ms knocks it down in 6 of 6. What does this teach?skill
Correct
In a calm walk, the policy corrects with room to spare. A push demands a fast reaction, and that's where every millisecond counts. A robustness test needs a disturbance.Not this one
In a calm walk, the policy corrects with room to spare. A push demands a fast reaction, and that's where every millisecond counts. A robustness test needs a disturbance.Review: Measure the latency and count the fallsQuestion 2 of 3
MuJoCo 3.5+ models latency in the actuator itself. What does the kit do in addition at startup?
Correct
The history starts zeroed: without filling it, the first N ms send a zero target to the joints, a jolt that doesn't exist on the robot.Not this one
The history starts zeroed: without filling it, the first N ms send a zero target to the joints, a jolt that doesn't exist on the robot.Review: Three ways to model the actuatorQuestion 3 of 3
When is it worth training an actuator network instead of PD with latency?
Correct
The network learns the command → torque relationship from motor data. It needs that data; without it, PD with latency and sysid is the way.Not this one
The network learns the command → torque relationship from motor data. It needs that data; without it, PD with latency and sysid is the way.Review: Three ways to model the actuatorQuiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as done.Pick one answer. The explanation appears after you try.
07In three sentences
MeasurementA step on the knee with the robot hanging measures the latency: 22 ms.
EffectUnder a push, falls rise from 2/6 (0 ms) to 6/6 (40 ms).
FixSending kp divided by the identified value makes the "simulated real" walk without falling.