torcs-devel Mailing List for TORCS - The Open Racing Car Simulator

Hi Owen

Interesting finding, but I am not yet sure that the position is the 
problem (float has an accuracy of ~8 decimals under suitable conditions).

What happens after 95 seconds, before the steps are way below 1 cm, 
after that they get quickly bigger?

What does the plot exactly show, how did you calculate the result (it 
must be more than just breaking, otherwise the distance would just 
develop in one direction)?

Maybe there is a floating point unfriendly expression in TORCS which 
could be rewritten in a more suitable way, or maybe this applies for 
your distance calculation.

Could you explain the whole thing in more detail? Any special things, 
e.g. floating point compiler flags (usually intermediate results are in 
80 or more bits in the FPU, but if you choose SSE as example, you have 
just 32 bit for floats, this can make a significant difference under 
certain conditions).

Kind regards

Bernhard

On 12/03/2014 11:01 AM, Owen McAree wrote:
> Hello developers,
>
> Firstly I'd like to say thanks for creating such a great piece of
> software! I've just started a research job and will be using TORCS
> extensively to study the behavior of autonomous vehicles.
>
> I have written a small interface between TORCS and MATLAB such that I
> can I can prototype my algorithms quickly and I have come across a small
> problem. The 'tdble' type used by TORCS is defined as a float, which for
> the majority of parameters makes sense but when working with the vehicle
> positions I have found some numerical errors creeping in to the
> integration inside SimCarUpdatePos (in car.cpp). The issue comes about
> when two vehicles have very similar (but not identical) velocities,
> which when integrated with the position get truncated to be identical.
> This has the effect that computing the distance between two vehicles
> returns a constant value (as the velocities are being made identical). I
> have attached a plot illustrating the bug, it shows the computed
> distance between two vehicles (one following the other with near
> identical velocity) when the lead vehicle brakes.
>
> At first I thought I could simply change the definition of tdble to
> double and this would solve the problem, but it seems that tdble is not
> used consistently through the source, so this caused more problems! In
> the end I modified the structures a little to simply have the position
> values represented as double (in addition to the tdble values, for
> compatibility) and leave the rest untouched, this has solved the problem
> (but is not elegant!).
>
> I appreciate this is an edge case which will only crop up very rarely,
> but if you'd like me to share my changes with you (or if you have a more
> elegant solution!) please let me know.
>
> Best regards,
> Owen
>
> --
> Dr Owen McAree
> Research Fellow
> Department of Automatic Control and Systems Engineering
> University of Sheffield, S13JD
> T: 0114 222 5138
> E: o.m...@sh... <mailto:o.m...@sh...>
>
>
> ------------------------------------------------------------------------------
> Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server
> from Actuate! Instantly Supercharge Your Business Reports and Dashboards
> with Interactivity, Sharing, Native Excel Exports, App Integration & more
> Get technology previously reserved for billion-dollar corporations, FREE
> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk
>
>
>
> _______________________________________________
> Torcs-devel mailing list
> Tor...@li...
> https://lists.sourceforge.net/lists/listinfo/torcs-devel
>

Hello developers,

Firstly I'd like to say thanks for creating such a great piece of software!
I've just started a research job and will be using TORCS extensively to
study the behavior of autonomous vehicles.

I have written a small interface between TORCS and MATLAB such that I can I
can prototype my algorithms quickly and I have come across a small problem.
The 'tdble' type used by TORCS is defined as a float, which for the
majority of parameters makes sense but when working with the vehicle
positions I have found some numerical errors creeping in to the integration
inside SimCarUpdatePos (in car.cpp). The issue comes about when two
vehicles have very similar (but not identical) velocities, which when
integrated with the position get truncated to be identical. This has the
effect that computing the distance between two vehicles returns a constant
value (as the velocities are being made identical). I have attached a plot
illustrating the bug, it shows the computed distance between two vehicles
(one following the other with near identical velocity) when the lead
vehicle brakes.

At first I thought I could simply change the definition of tdble to double
and this would solve the problem, but it seems that tdble is not used
consistently through the source, so this caused more problems! In the end I
modified the structures a little to simply have the position values
represented as double (in addition to the tdble values, for compatibility)
and leave the rest untouched, this has solved the problem (but is not
elegant!).

I appreciate this is an edge case which will only crop up very rarely, but
if you'd like me to share my changes with you (or if you have a more
elegant solution!) please let me know.

Best regards,
Owen

-- 
Dr Owen McAree
Research Fellow
Department of Automatic Control and Systems Engineering
University of Sheffield, S13JD
T: 0114 222 5138
E: o.m...@sh...