One PPO policy for Go1JoystickFlatTerrain, exported to ONNX nine times at
different weight precisions, plus the scene it was trained in. The point of the
set is the comparison: how far precision can be reduced before a closed-loop
controller stops working, and where that boundary actually sits.
100 episodes x 5 fixed seeds, 1000-step horizon, deterministic policy (mean
action, no sampling). Return deltas are paired bootstrap 95% intervals over
10 000 resamples. P50 is the perturbation magnitude at which the success rate
crosses 50%; on friction_scale, swept downward, a larger number is worse.
scheme
bits/weight
mean return
Δreturn vs fp32 (95% CI)
P50 friction
P50 obs noise
fp32
32.00
31.65
—
0.189
0.142
fp16
16.00
31.65
no measurable loss
0.192
0.143
int8-tensor
8.00
31.68
no measurable loss
0.200
0.142
int8-channel
8.00
31.66
no measurable loss
0.205
0.142
int4-channel
4.00
30.91
-2.336% [-2.79%, -1.95%]
0.241
0.127
int4-group32
4.00
31.42
-0.728% [-0.88%, -0.45%]
0.207
0.140
ternary
1.58
2.95
-90.672% [-92.14%, -89.13%]
1.000
0.000
mixed-head-fp16
4.19
31.36
-0.902% [-1.26%, -0.70%]
0.198
0.136
int8-act
8.00
31.65
no measurable loss
0.193
0.142
What the numbers say
int8 is free. Every int8 variant, including one that quantizes activations
as well as weights, is statistically indistinguishable from fp32 on return and
on all five robustness axes.
The boundary is between int8 and int4, and grouping matters more than bits.int4-channel loses only 2.3% of return on flat ground yet its friction and
observation-noise frontiers separate from fp32 with non-overlapping intervals.
int4-group32, at the same 4.00 bits, does not.
Quantizing the observation-normalization statistics is catastrophic. They
are 0.05% of the parameters. Quantized with the same scheme as the weights,
every int8 variant becomes measurably lossy and every 4-bit variant stops
producing finite actions at all, because a strictly positive scale vector
quantized symmetrically rounds entries to zero and the policy divides by them.
Intended use and limitations
These are research artifacts for studying quantization of closed-loop control.
They are not tuned for deployment on hardware and have never been run on a
physical robot.
Quantization is simulated. Weights are rounded to the target grid and
stored back as float32, so every ONNX file is the same size and none of them
runs faster. The benchmark measures what precision does to control, not what
it does to storage or throughput. A deployment would need packed kernels, and
the accuracy results here would carry over while the timing results would not.
One training seed. Precision effects cannot be fully separated from the
luck of a single checkpoint.
One task, flat terrain.Go1JoystickFlatTerrain only. The headline
hypothesis (that open-loop action error mispredicts closed-loop performance)
is not supported on this task, and that negative result is reported with the
same prominence as the positive ones.
Browser physics differs from training physics. The demo runs MuJoCo 3.3.8
compiled to WebAssembly; training used MJX with MuJoCo 3.10.0. Measured
divergence and the reasoning behind it are in the repository's methodology.
Training
Environment steps executed
206,438,400
Wall clock
140.0 min
Parallel environments
8,192
Peak VRAM
2225 MiB
Final training reward
31.772 ± 1.810
Weight extraction vs Brax
4.47e-06
Hardware
NVIDIA RTX 3050 Laptop, 4096 MiB
Every ONNX graph carries the observation normalization inside it, so the input
is the raw 48-dim observation and the output is the tanh-squashed action.
