success in recovering both previously unknown
cave bear and known Neanderthal genomic
sequences using direct genomic selection indicates
that this is a feasible strategy for purifying specific
cloned Neanderthal sequences out of a high
background of Neanderthal and contaminating
microbial DNA. This raises the possibility that,
should multiple Neanderthal metagenomic libra-
ries be constructed from independent samples,
direct selection could be used to recover Neander-
thal sequences from several individuals to obtain
and confirm important human-specific and Nean-
derthal-specific substitutions.
Conclusions. The current state of our knowl-
edge concerning Neanderthals and their relationship
to modern humans is largely inference and speculation
based on archaeological data and a limited number of
hominid remains. In this study, we have demonstrated
that Neanderthal genomic sequences can be recovered
using a metagenomic library-based approach and that
specific Neanderthal sequences can be obtained from
such libraries by direct selection. Our study thus pro-
vides a framework for the rapid recovery of Nean-
derthal sequences of interest from multiple
independent specimens, without the need for whole-
genome resequencing. Such a collection of targeted
Neanderthal sequences would be of immense value
for understanding human and Neanderthal biology
and evolution. Future Neanderthal genomic studies,
including targeted and whole-genome shotgun
sequencing, will provide insight into the profound
phenotypic divergence of humans both from the great
apes and from our extinct hominid relatives, and will
allow us to explore aspects of Neanderthal biology not
evident from artifacts and fossils.
References and Notes
1. P. Mellars, Nature 432, 461 (2004).
2. F.H.Smith,E.Trinkaus,P.B.Pettitt,I.Karavanic,M.Paunovic,
Proc. Natl. Acad. Sci. U.S.A. 96, 12281 (1999).
3. M. Krings et al., Cell 90, 19 (1997).
4. M. Krings et al., Proc. Natl. Acad. Sci. U.S.A. 96, 5581
(1999).
5. S. Pääbo et al., Annu. Rev. Genet. 38, 645 (2004).
6. D. Serre et al., PLoS Biol. 2, e57 (2004).
7. M. Currat, L. Excoffier, PLoS Biol. 2, e421 (2004).
8. S. G. Tringe et al., Science 308, 554 (2005).
9. S. G. Tringe, E. M. Rubin, Nat. Rev. Genet. 6, 805 (2005).
10. J. P. Noonan et al., Science 309, 597 (2005).
11. M. Margulies et al., Nature 437, 376 (2005).
12. H. N. Poinar et al., Science 311, 392 (2006).
13. Materials and methods are available as supporting
material on Science Online.
14. S. F. Altschul et al., Nucleic Acids Res. 25, 3389 (1997).
15. Chimpanzee Sequencing and Analysis Consortium, Nature
437, 69 (2005).
16. M. Hofreiter et al., Nucleic Acids Res. 29, 4793 (2001).
17. S. Kumar, A. Filipski, V. Swarna, A. Walker, S. B. Hedges,
Proc. Natl. Acad. Sci. U.S.A. 102, 18842 (2005).
18. N. Patterson, D. Richter, S. Gnerre, E. Lander, D. Reich,
Nature, in press; published online 17 May 2006
(10.1038/nature04789).
19. The International HapMap Consortium et al., Nature
437, 1299 (2005).
20. B. F. Voight et al., Proc. Natl. Acad. Sci. U.S.A. 102,
18508 (2005).
21. A. M. Adams, R. R. Hudson, Genetics 168, 1699 (2004).
22. I. McDougall et al., Nature 433, 733 (2005).
23. V. Plagnol, J. D. Wall, PLoS Genet., in press (110.1371/
journal.pgen.0020105.eor).
24. S. Bashiardes et al., Nat. Methods 2, 63 (2005).
25. P. Mellars, Nature 439, 931 (2006).
26. Neanderthal sequences reported in this study have been
deposited in GenBank under accession numbers DX935178
to DX936503.We thank E. Green, M. Lovett, andmembers of
the Rubin, Pääbo, and Pritchard laboratories for insightful
discussions and support. J.P.N. was supported by NIH
National Research Service Award fellowship 1-F32-
GM074367. G.C. and S.K. were supported by grant R01
HG002772-1 (NIH) to J.K.P. This work was supported by
grant HL066681, NIH Programs for Genomic Applications,
funded by the National Heart, Lung and Blood Institute; and
by the Director, Office of Science, Office of Basic Energy
Sciences, of the U.S. Department of Energy under contract
number DE-AC02-05CH11231.
Supporting Online Material
www.sciencemag.org/cgi/content/full/314/5802/1113/DC1
Materials and Methods
Figs. S1 to S6
Tables S1 to S12
References
16 June 2006; accepted 17 August 2006
10.1126/science.1131412
REPORTS
Resilient Machines Through
Continuous Self-Modeling
Josh Bongard,
1
*† Victor Zykov,
1
Hod Lipson
1,2
Animals sustain the ability to operate after injury by creating qualitatively different compensatory
behaviors. Although such robustness would be desirable in engineered systems, most machines fail
in the face of unexpected damage. We describe a robot that can recover from such change
autonomously, through continuous self-modeling. A four-legged machine uses actuation-sensation
relationships to indirectly infer its own structure, and it then uses this self-model to generate
forward locomotion. When a leg part is removed, it adapts the self-models, leading to the
generation of alternative gaits. This concept may help develop more robust machines and shed
light on self-modeling in animals.
R
obotic systems are of growing interest
because of their many practical applica-
tions as well as their ability to help
understand human and animal behavior (1–3),
cognition (4–6), and physical performance (7).
Although industrial robots have long been used
for repetitive tasks in structured environments,
one of the long-standing challenges is achieving
robust performance under uncertainty (8). Most
robotic systems use a manually constructed
mathematical model that captures the robot’s
dynamicsandisthenusedtoplanactions(9).
Although some parametric identification methods
exist for automatically improving these models
(10–12), making accurate models is difficult for
complex machines, especially when trying to
account for possible topological changes to the
body, such as changes resulting from damage.
Fig. 7. Recovery of Neanderthal genomic sequences from library NE1 by direct genomic selection.
1
Mechanical and Aerospace Engineering, Cornell Univer-
sity, Ithaca, NY 14853, USA.
2
Computing and Information
Science, Cornell University, Ithaca, NY 14853, USA.
*Present address: Department of Computer Science,
University of Vermont, Burlington, VT 05405, USA.
†
To whom correspondence should be addressed. E-mail:
josh.bongard@uvm.edu
17 NOVEMBER 2006 VOL 314 SCIENCE www.sciencemag.org1118

Although much progress has been made in
allowing robotic systems to model their environ-
ment autonomously (8), relatively little is known
about how a robot can learn its own morphology,
whichcannotbeinferredbydirectobservationor
retrieved from a database of past experiences (13).
Without internal models, robotic systems can auton-
omously synthesize increasingly complex behaviors
(6, 14–16)orrecoverfromdamage(17) through
physical trial and error, but this requires hundreds or
thousands of tests on the physical machine and is
generally too slow, energetically costly, or risky.
Here, we describe an active process that allows
a machine to sustain performance through an
autonomous and continuous process of self-
modeling. A robot is able to indirectly infer its
own morphology through self-directed exploration
and then use the resulting self-models to synthesize
new behaviors. If the robot’s topology unexpect-
edly changes, the same process restructures its
internal self-models, leading to the generation of
qualitatively different, compensatory behavior. In
essence, the process enables the robot to continu-
ously diagnose and recover from damage. Unlike
other approaches to damage recovery, the concept
introduced here does not presuppose built-in
redundancy (18, 19), dedicated sensor arrays, or
contingency plans designed for anticipated failures
(20). Instead, our approach is based on the concept
of multiple competing internal models and gener-
ation of actions to maximize disagreement
between predictions of these models.
The process is composed of three algorithmic
components that are executed continuously by the
physical robot while moving or at rest (Fig. 1):
Modeling, testing, and prediction. Initially, the
robot performs an arbitrary motor action and
records the resulting sensory data (Fig. 1A). The
model synthesis component (Fig. 1B) then syn-
thesizes a set of 15 candidate self-models using
stochastic optimization to explain the observed
sensory-actuation causal relationship. The action
synthesis component (Fig. 1C) then uses these
models to find a new action most likely to elicit the
most information from the robot. This is
accomplished by searching for the actuation pattern
that, when executed on each of the candidate self-
models, causes the most disagreement across the
predicted sensor signals (21–24). This new action
is performed by the physical robot (Fig. 1A), and
the model synthesis component now reiterates with
more available information for assessing model
quality. After 16 cycles of this process have
terminated, the most accurate model is used by
the behavior synthesis component to create a
desired behavior (Fig. 1D) that can then be
executed by the robot (Fig. 1E). If the robot detects
unexpected sensor-motor patterns or an external
signal as a result of unanticipated morphological
change, the robot reinitiates the alternating cycle of
modeling and exploratory actions to produce new
models reflecting the change. The new most
accurate model is now used to generate a new,
compensating behavior to recover functionality. A
complete sample experiment is shown in Fig. 2.
We tested the proposed process on a four-
legged physical robot that had eight motorized
joints, eight joint angle sensors, and two tilt sensors.
The space of possible models comprised any planar
topological arrangement of eight limbs, including
chains and trees (for examples, see Figs. 1 and 2).
After damage occurs, the space of topologies is
fixed to the previously inferred morphology, but the
size of the limbs can be scaled (Fig. 2, N and O).
The space of possible actions comprised desired
angles that the motors were commanded to reach
(25). Many other self-model representations could
replace the explicit simulations used here, such as
artificial neural or Bayesian networks, and other
sensory modalities could be exploited, such as
pressure and acceleration (here the joint angle
sensors were used only to verify achievement of
desired angles, and orientation of the main body
was used only for self-model synthesis). None-
theless, the use of implicit representations such as
artificial neural networks—although more biologi-
cally plausible than explicit simulation—would
make the validation of our theory more challenging,
because it would be difficult to assess the
correctness of the model (which can be done by
visual inspection for explicit simulations). More
important, without an explicit representation, it is
difficult to reward a model for a task such as
forward locomotion (which requires predictions
about forward displacement) when the model can
only predict orientation data.
The proposed process was compared with two
baseline algorithms, both of which use random
rather than self-model–driven data acquisition. All
three algorithm variants used a similar amount of
computational effort (~250,000 internal model
simulations) and the same number (16) of physical
actions (Table 1). In the first baseline algorithm, 16
random actions were executed by the physical robot
(Fig. 1A), and the resulting data were supplied to
the model synthesis component for batch training
(Fig.1B).Inthesecondbaseline algorithm, the
action synthesis component output a random action,
rather than searching for one that created dis-
agreement among competing candidate self-
models. The actions associated with Fig. 1, A to C,
Fig. 1. Outline of the algorithm. The robot continuously cycles through action execution. (A and B)
Self-model synthesis. The robot physically performs an action (A). Initially, this action is random;
later, it is the best action found in (C). The robot then generates several self-models to match
sensor data collected while performing previous actions (B). It does not know which model is
correct. (C) Exploratory action synthesis. The robot generates several possible actions that
disambiguate competing self-models. (D) Target behavior synthesis. After several cycles of (A) to
(C), the currently best model is used to generate locomotion sequences through optimization. (E)
The best locomotion sequence is executed by the physical device. (F) The cycle continues at step (B)
to further refine models or at step (D) to create new behaviors.
www.sciencemag.org SCIENCE VOL 314 17 NOVEMBER 2006 1119
REPORTS