What Went Right
Perhaps Naughty Dog's most important achievement is making large-scale
games and shipping them on time, with at most a small amount of slip.
This is an almost unheard of combination in the industry, and although
there is a certain amount of luck involved, there are valid reasons to
explain how Naughty Dog has managed to achieve this time and again.
Experience will tell you it's impossible to predict the precise details
of what will be worked on more than a month or two in advance of doing
it, yet many companies fall into the trap of trying to maintain a highly
detailed schedule that tries to look too far into the future. What can't
be effectively worked into these rigid schedules is time lost to debugging,
design changes, over-optimism, illness, meetings, new ideas, and myriad
other unpredictable surprises.
At Naughty Dog, we prefer a much more flexible, macro-level scheduling
scheme, with milestone accomplishments to be achieved by certain dates.
The schedule only becomes detailed for tasks that will be tackled in the
near future. For example, a certain level will be scheduled to have its
background modeled by a certain date. If the milestone is missed, then
the team makes an analysis as to why the milestone wasn't achieved and
changes plans accordingly: the background may be reduced in size, a future
task of that artist may be given to another artist to free up more time,
the artist may receive guidance on how to model more productively, or
some future task may be eliminated.
In the case of Jak & Daxter, we used the knowledge we'd gained
from creating the Crash Bandicoot games to help estimate how long it should
take to model a level. As we modeled a few levels, however, we soon realized
that our original estimates were far too short, and so we took appropriate
actions. If we had attempted to maintain a long-term, rigidly detailed
schedule, we would have spent a lot of time trying to update something
that was highly inaccurate. Beyond this being a waste of time, the constant
rescheduling could have had a demoralizing effect on the team.
localization techniques. We knew from the start that we were going
to sell Jak & Daxter into many territories around the world, so we
knew we would face many localization issues, such as PAL-versus-NTSC,
translations, and audio in multiple languages. Careful structuring of
our game code and data allowed us to localize to a particular territory
by swapping a few data files. This meant we only had to debug one executable
and that we had concurrent development of all localized versions of the
All of our animation playback code was written so that it could automatically
step animations at a rate of 1.2 (60fps/50fps) when playing in PAL. We
also used a standardized number of units per second so that we could relate
the amount of time elapsed in a game frame to our measure of units per
second. Once everything was nice and consistent, then timing-related code
no longer had to be concerned with the differences between PAL and NTSC.
Physics calculations were another issue. If a ball's motion while being
dropped is computed by adding a gravitational force to the ball's velocity
every frame, then after one second the ball's velocity has been accelerated
by gravity 60 times in NTSC but only 50 times in PAL. This discrepancy
was big enough to become problematic between the two modes. To correct
this problem, we made sure that all of our physics computations were done
using seconds, and then we converted the velocity-per-second into velocity-per-game-frame
before adding the velocity to the translation.
world, grand vistas, and no load times. We knew very early on in the
development of Jak & Daxter that we wanted to immerse the player
within one large expansive world. We didn't want to stall the game with
loads between the various areas of that world.
Jak & Daxter's designers had to overcome many obstacles to
achieve our open environments. They had to lay out the levels of the world
carefully so that levels could be moved in and out of memory without stalling
gameplay or causing ugly visual popping. They also had to create challenges
that would engage the player and maintain the player's interest, even
though the player could roam freely around the world. And they had to
tune the challenges so that the difficulty ramped up appropriately, without
giving players the impression that they were being overly directed.
The programmers had to create tools to process interconnected levels containing
millions of polygons and create the fast game code that could render the
highly detailed world. We developed several complex level-of-detail (LOD)
schemes, with different schemes used for different types of things (creatures
versus background), and different schemes used at different distances,
such as simplified models used to represent faraway backgrounds, and flats
used to represent distant geometry. At the heart of our LOD system was
our proprietary mesh tessellation/reduction scheme, which we originally
developed for Crash Team Racing and radically enhanced for Jak &
The artists had the burden of generating the enormous amount of content
for these environments. Their task was complicated by the very specialized
construction rules they had to follow to support our various renderers.
Support tools and plug-ins were created to help the artists, but we relied
on the art staff to overcome many difficulties.
Camera control. From the initial stages of Jak & Daxter,
we looked at the various camera schemes used in other games and came to
the depressing conclusion that all existing camera schemes had serious
issues. We suspected that making a well-behaved camera might be an unsolvable
3D problem: How could one possibly create a camera that would maneuver
through a complex 3D world while behaving both unobtrusively and intelligently?
Only fools would believe that all problems have a solution, so, like idiots,
we decided to give it a try. The resulting camera behaved extremely well,
and although it had its limitations, it proved the problem does indeed
have a solution. Jak can jump through trees and bushes, duck under archways,
run between scaffolding, scale down cliffs, and hide behind rocks, all
with the camera unobtrusively keeping the action in view.
We wanted the player to be able to control the camera, but we did not
want to force the player to do so. Players can use the second joystick
to maneuver the camera(rotating the camera or moving it closer to or farther
from Jak), but we were concerned that some people may not want to manipulate
the camera, and others, such as children, may not have the required sophistication
or coordination. Therefore, we worked very hard at making the camera do
a reasonable job of showing players what they needed to see in order to
complete the various challenges. We accomplished this through a combination
of camera volumes with specially tuned camera parameters and specialized
camera modes for difficult situations. Also, creatures could send messages
to the camera in order to help the camera better show the action.
This may sound funny, but an important feature of the camera was that
it didn't make people sick. This has been a serious problem that has plagued
cameras in other games. We spent a bit of time analyzing why people got
sick, and we tuned the camera so that it reduced the rotational and extraneous
movement that contributed to the problem.
Perhaps the greatest success of the camera is that everyone seems to like
it. We consider that a major accomplishment, given the difficulty of the
task of creating it.
rules! Practically all of the run-time code (approximately half a
million lines of source code) was written in GOAL (Game Object Assembly
Lisp), Naughty Dog's own internally developed language, which was based
on the Lisp programming language. Before you dismiss us as crazy, consider
the many advantages of having a custom compiler.
Lisp has a very consistent, small set of syntactic rules involving the
construction and evaluation of lists. Lists that represent code are executed
by evaluating the items that are in the list; if the head of the list
is a function (or some other action), you could think of the other items
in the list as being the parameters to that function. This simplicity
of the Lisp syntax makes it trivial to create powerful macros that would
be difficult or impossible to implement using C++.
Writing macros, however, is not enough justification for writing a compiler;
there were features we felt we couldn't achieve without a custom compiler.
GOAL code, for example, can be executed at a listener prompt while the
game is running. Not only can numbers be viewed and tweaked, code itself
can be compiled and downloaded without interrupting or restarting the
game. This allowed the rapid tuning and debugging, since the effects of
modifying functions and data structures could be viewed instantaneously.
We wanted creatures to use nonpreemptive cooperative multi-tasking, a
fancy way of saying that we wanted a creature to be able to execute code
for a while, then "suspend" and allow other code to execute.
The advantage of implementing the multi-tasking scheme using our own language
was that suspend instructions could be inserted within a creature's code,
and state could be automatically preserved around the suspend. Consider
the following small snippet of GOAL code:
(dotimes (ii (num-frames
(set! frame-num ii)
has been simplified to make a point, so pretend that it uses a counter
called ii to loop over the number of frames in an animation called idle.
Each time through the loop the animation frame is set to the value of
ii, and the code is suspended. Note that the value of ii (as well as any
other local variables) is automatically preserved across the suspend.
In practice, the preceding code would have been encapsulated into a macro
;; Put code executed for each time..
;; through the loop here.
There are other major compiler advantages: a unified set of assembly op-codes
consistent across all five processors of the Playstation 2, register coloring
when writing assembly code, and the ability to intermix assembly instructions
seamlessly with higher-level code. Outer loops could be written as "slower"
higher-level code, while inner loops could be optimized assembly.