Limits...
Distributed recurrent neural forward models with synaptic adaptation and CPG-based control for complex behaviors of walking robots.

Dasgupta S, Goldschmidt D, Wörgötter F, Manoonpong P - Front Neurorobot (2015)

Bottom Line: Biological study has revealed that such complex behaviors are a result of a combination of biomechanics and neural mechanism thus representing the true nature of embodied interactions.Inspired by these findings, we present here, an artificial bio-inspired walking system which effectively combines biomechanics (in terms of the body and leg structures) with the underlying neural mechanisms.Furthermore, we demonstrate that the newly developed recurrent network based approach to online forward models outperforms the adaptive neuron forward models, which have hitherto been the state of the art, to model a subset of similar walking behaviors in walking robots.

View Article: PubMed Central - PubMed

Affiliation: Institute for Physics - Biophysics, George-August-University Göttingen, Germany ; Bernstein Center for Computational Neuroscience, George-August-University Göttingen, Germany ; Laboratory for Neural Computation and Adaptation, Riken Brain Science Institute Saitama, Japan.

ABSTRACT
Walking animals, like stick insects, cockroaches or ants, demonstrate a fascinating range of locomotive abilities and complex behaviors. The locomotive behaviors can consist of a variety of walking patterns along with adaptation that allow the animals to deal with changes in environmental conditions, like uneven terrains, gaps, obstacles etc. Biological study has revealed that such complex behaviors are a result of a combination of biomechanics and neural mechanism thus representing the true nature of embodied interactions. While the biomechanics helps maintain flexibility and sustain a variety of movements, the neural mechanisms generate movements while making appropriate predictions crucial for achieving adaptation. Such predictions or planning ahead can be achieved by way of internal models that are grounded in the overall behavior of the animal. Inspired by these findings, we present here, an artificial bio-inspired walking system which effectively combines biomechanics (in terms of the body and leg structures) with the underlying neural mechanisms. The neural mechanisms consist of (1) central pattern generator based control for generating basic rhythmic patterns and coordinated movements, (2) distributed (at each leg) recurrent neural network based adaptive forward models with efference copies as internal models for sensory predictions and instantaneous state estimations, and (3) searching and elevation control for adapting the movement of an individual leg to deal with different environmental conditions. Using simulations we show that this bio-inspired approach with adaptive internal models allows the walking robot to perform complex locomotive behaviors as observed in insects, including walking on undulated terrains, crossing large gaps, leg damage adaptations, as well as climbing over high obstacles. Furthermore, we demonstrate that the newly developed recurrent network based approach to online forward models outperforms the adaptive neuron forward models, which have hitherto been the state of the art, to model a subset of similar walking behaviors in walking robots.

No MeSH data available.


Related in: MedlinePlus

Real-time data of walking and crossing multiple gaps using the forward model predictions. (A) The accumulated error (black line) and the maximum accumulated error value at the end of each stance phase (red line) of the right front leg (R1). The accumulated error is reset to zero every swing phase. (B) The backbone joint (BJ) angle during walking and gap crossing. The BJ stays at the normal position (−2°) during normal walking. On encountering a gap (15cm), it leans upwards in a step like fashion and then finally bent downwards in order to cross the gap. This procedure is repeated for the second gap (11cm), however with different degree of elevations. (C–E) The TC-, CTr-, and FTi-joint angles of right front leg R1 during normal walking and gap crossing. The joint adaptation was controlled by the maximum accumulated error value of the previous step (red line). Below pictures show snap shots of the locomotion of AMOS II during the experiment. Note that one time step is ≈0.037 s. For further details interested readers are recommended to see the experiment Supplementary Videos 1, 2.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4585172&req=5

Figure 7: Real-time data of walking and crossing multiple gaps using the forward model predictions. (A) The accumulated error (black line) and the maximum accumulated error value at the end of each stance phase (red line) of the right front leg (R1). The accumulated error is reset to zero every swing phase. (B) The backbone joint (BJ) angle during walking and gap crossing. The BJ stays at the normal position (−2°) during normal walking. On encountering a gap (15cm), it leans upwards in a step like fashion and then finally bent downwards in order to cross the gap. This procedure is repeated for the second gap (11cm), however with different degree of elevations. (C–E) The TC-, CTr-, and FTi-joint angles of right front leg R1 during normal walking and gap crossing. The joint adaptation was controlled by the maximum accumulated error value of the previous step (red line). Below pictures show snap shots of the locomotion of AMOS II during the experiment. Note that one time step is ≈0.037 s. For further details interested readers are recommended to see the experiment Supplementary Videos 1, 2.

Mentions: Thus, we use the positive value for searching control (Figure 6D, above). This is then accumulated through a single recurrent neuron S with a linear transfer function and is always reset to 0.0 at the beginning of swing phase. Similarly, the negative value is used for elevation control (Figure 6D, below). The value is also accumulated through a recurrent neuron E with a linear transfer function. These accumulated errors (Figure 6C) thus allow the robot leg to be either elevated (on hitting an obstacle) or searching for a foothold during the swing and stance phases, respectively (see Manoonpong et al., 2013, for more details of the searching and elevation control). As depicted in Figures 6A,B, while walking on a rough terrain (in this case with tetrapod walking gait), the currently recorded sensory feedback or foot contact sensor reading differs considerably from the reservoir predicted signal. As a result, there is a high accumulation of error between each swing or stance phase (Figure 6C). It should be noted that the initial (≈50 time steps) abruptly high amplitude signal observed in the reservoir forward model prediction, is caused due to the transient recovery time needed by reservoir readout neurons to settle to the exact learned patterns. This is overcome within the next few time steps and RF predicted FC signal continues to occur in a robust manner. The accumulated error causes the corresponding leg action control mechanism to kick in and the robot successfully navigates out of the rough terrain (after ≈4000 time steps). Once the robot moves into the flat terrain, the reservoir predicted foot contact signal matches almost perfectly with the actual sensory feedback. As a result, the accumulated error becomes zero and normal walking without any additional searching or elevation control mechanisms, can continue. In essence based on the reservoir forward models, while traversing from the uneven terrain (Figure 6, inset 1–4) to the flat terrain (Figure 6, inset 5), the robot can adapt its legs individually to deal with the change of terrain. That is, it depressed its leg and extended its tibia to search for a foothold when loosing ground contact during the stance phase. Losing ground contact information is detected by a significant change of the accumulated errors (Figure 6C). In case of both walking on uneven terrain and climbing, this accumulated error causes shifting of the CTr- and FTi-joints causing the respective leg to search for a foothold. However, in the specific case of crossing a gap (Figure 7), we use the accumulated error in order to control tilting of the backbone joint (BJ) and shifting of the TC- and FTi-joints such that the front legs can be extended forward continuously till the robot can find a foothold. In addition to this leg joint control, reactive backbone joint control using the additional ultrasonic sensors in front of the robot can also be used to learn to lean up the BJ for climbing over obstacles (this has been previously successfully applied using classical conditioning based learning in Goldschmidt et al. (2014) and as such not discussed here).


Distributed recurrent neural forward models with synaptic adaptation and CPG-based control for complex behaviors of walking robots.

Dasgupta S, Goldschmidt D, Wörgötter F, Manoonpong P - Front Neurorobot (2015)

Real-time data of walking and crossing multiple gaps using the forward model predictions. (A) The accumulated error (black line) and the maximum accumulated error value at the end of each stance phase (red line) of the right front leg (R1). The accumulated error is reset to zero every swing phase. (B) The backbone joint (BJ) angle during walking and gap crossing. The BJ stays at the normal position (−2°) during normal walking. On encountering a gap (15cm), it leans upwards in a step like fashion and then finally bent downwards in order to cross the gap. This procedure is repeated for the second gap (11cm), however with different degree of elevations. (C–E) The TC-, CTr-, and FTi-joint angles of right front leg R1 during normal walking and gap crossing. The joint adaptation was controlled by the maximum accumulated error value of the previous step (red line). Below pictures show snap shots of the locomotion of AMOS II during the experiment. Note that one time step is ≈0.037 s. For further details interested readers are recommended to see the experiment Supplementary Videos 1, 2.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4585172&req=5

Figure 7: Real-time data of walking and crossing multiple gaps using the forward model predictions. (A) The accumulated error (black line) and the maximum accumulated error value at the end of each stance phase (red line) of the right front leg (R1). The accumulated error is reset to zero every swing phase. (B) The backbone joint (BJ) angle during walking and gap crossing. The BJ stays at the normal position (−2°) during normal walking. On encountering a gap (15cm), it leans upwards in a step like fashion and then finally bent downwards in order to cross the gap. This procedure is repeated for the second gap (11cm), however with different degree of elevations. (C–E) The TC-, CTr-, and FTi-joint angles of right front leg R1 during normal walking and gap crossing. The joint adaptation was controlled by the maximum accumulated error value of the previous step (red line). Below pictures show snap shots of the locomotion of AMOS II during the experiment. Note that one time step is ≈0.037 s. For further details interested readers are recommended to see the experiment Supplementary Videos 1, 2.
Mentions: Thus, we use the positive value for searching control (Figure 6D, above). This is then accumulated through a single recurrent neuron S with a linear transfer function and is always reset to 0.0 at the beginning of swing phase. Similarly, the negative value is used for elevation control (Figure 6D, below). The value is also accumulated through a recurrent neuron E with a linear transfer function. These accumulated errors (Figure 6C) thus allow the robot leg to be either elevated (on hitting an obstacle) or searching for a foothold during the swing and stance phases, respectively (see Manoonpong et al., 2013, for more details of the searching and elevation control). As depicted in Figures 6A,B, while walking on a rough terrain (in this case with tetrapod walking gait), the currently recorded sensory feedback or foot contact sensor reading differs considerably from the reservoir predicted signal. As a result, there is a high accumulation of error between each swing or stance phase (Figure 6C). It should be noted that the initial (≈50 time steps) abruptly high amplitude signal observed in the reservoir forward model prediction, is caused due to the transient recovery time needed by reservoir readout neurons to settle to the exact learned patterns. This is overcome within the next few time steps and RF predicted FC signal continues to occur in a robust manner. The accumulated error causes the corresponding leg action control mechanism to kick in and the robot successfully navigates out of the rough terrain (after ≈4000 time steps). Once the robot moves into the flat terrain, the reservoir predicted foot contact signal matches almost perfectly with the actual sensory feedback. As a result, the accumulated error becomes zero and normal walking without any additional searching or elevation control mechanisms, can continue. In essence based on the reservoir forward models, while traversing from the uneven terrain (Figure 6, inset 1–4) to the flat terrain (Figure 6, inset 5), the robot can adapt its legs individually to deal with the change of terrain. That is, it depressed its leg and extended its tibia to search for a foothold when loosing ground contact during the stance phase. Losing ground contact information is detected by a significant change of the accumulated errors (Figure 6C). In case of both walking on uneven terrain and climbing, this accumulated error causes shifting of the CTr- and FTi-joints causing the respective leg to search for a foothold. However, in the specific case of crossing a gap (Figure 7), we use the accumulated error in order to control tilting of the backbone joint (BJ) and shifting of the TC- and FTi-joints such that the front legs can be extended forward continuously till the robot can find a foothold. In addition to this leg joint control, reactive backbone joint control using the additional ultrasonic sensors in front of the robot can also be used to learn to lean up the BJ for climbing over obstacles (this has been previously successfully applied using classical conditioning based learning in Goldschmidt et al. (2014) and as such not discussed here).

Bottom Line: Biological study has revealed that such complex behaviors are a result of a combination of biomechanics and neural mechanism thus representing the true nature of embodied interactions.Inspired by these findings, we present here, an artificial bio-inspired walking system which effectively combines biomechanics (in terms of the body and leg structures) with the underlying neural mechanisms.Furthermore, we demonstrate that the newly developed recurrent network based approach to online forward models outperforms the adaptive neuron forward models, which have hitherto been the state of the art, to model a subset of similar walking behaviors in walking robots.

View Article: PubMed Central - PubMed

Affiliation: Institute for Physics - Biophysics, George-August-University Göttingen, Germany ; Bernstein Center for Computational Neuroscience, George-August-University Göttingen, Germany ; Laboratory for Neural Computation and Adaptation, Riken Brain Science Institute Saitama, Japan.

ABSTRACT
Walking animals, like stick insects, cockroaches or ants, demonstrate a fascinating range of locomotive abilities and complex behaviors. The locomotive behaviors can consist of a variety of walking patterns along with adaptation that allow the animals to deal with changes in environmental conditions, like uneven terrains, gaps, obstacles etc. Biological study has revealed that such complex behaviors are a result of a combination of biomechanics and neural mechanism thus representing the true nature of embodied interactions. While the biomechanics helps maintain flexibility and sustain a variety of movements, the neural mechanisms generate movements while making appropriate predictions crucial for achieving adaptation. Such predictions or planning ahead can be achieved by way of internal models that are grounded in the overall behavior of the animal. Inspired by these findings, we present here, an artificial bio-inspired walking system which effectively combines biomechanics (in terms of the body and leg structures) with the underlying neural mechanisms. The neural mechanisms consist of (1) central pattern generator based control for generating basic rhythmic patterns and coordinated movements, (2) distributed (at each leg) recurrent neural network based adaptive forward models with efference copies as internal models for sensory predictions and instantaneous state estimations, and (3) searching and elevation control for adapting the movement of an individual leg to deal with different environmental conditions. Using simulations we show that this bio-inspired approach with adaptive internal models allows the walking robot to perform complex locomotive behaviors as observed in insects, including walking on undulated terrains, crossing large gaps, leg damage adaptations, as well as climbing over high obstacles. Furthermore, we demonstrate that the newly developed recurrent network based approach to online forward models outperforms the adaptive neuron forward models, which have hitherto been the state of the art, to model a subset of similar walking behaviors in walking robots.

No MeSH data available.


Related in: MedlinePlus