Deep Eigen

Making cutting-edge education sustainable.

[Sanjeev Sharma @sanjeevs_iitr]

Level-5 autonomous driving requires autonomous vehicles to negotiate not just unpredictable, and unforeseeable situations on road, but also to negotiate extremely tight spaces. One of the hottest test beds for this scenario is Indian traffic, where one may be required to negotiate bumper to bumper traffic, often obstacles coming within a meter (or a few centimeters) of each other.

In this demo we showcase our #autonomousdriving technology enabling our Deep Xplorer autonomous vehicle negotiating tightly placed obstacles as one of the first tests in this direction. The vehicle was tasked with negotiating a narrow road with static obstacles placed in relative proximities. The vehicle was required to perform lane change maneuvers, obstacles negotiation, and braking all on its own -- one can notice that between 1:10 and 1:22 mark the vehicle didn't perform unnecessary lane-change maneuver -- an expected ideal behavior in this scenario.

One of the major points was that this was done without any human presence in our vehicle, on a very narrow road with public vehicles parked throughout the road segments. Our next test in this endeavor would be to allow negotiation dynamic obstacles moving in proximity to, and overtaking, our #autonomousvehicles. In around a month from now, we will be testing the same in generalized uncontrolled traffic.

#reinforcementlearning #deeplearning @swaayatt

[Deep Eigen @DeepEigen]

Autonomous dodging off-roads -- without any human presence by @swaayatt @sanjeevs_iitr in their #AutonomousVehicle Deep Xplorer, pushing the limits of what is possible in off-road #AutonomousDriving.

#ReinforcementLearning #DeepLearning

[Swaayatt Robots @swaayatt]

Autonomous dodging off-roads -- without any human presence. This is a demonstration of pushing the limits of what is possible in off-road #AutonomousDriving.

In this demo, autonomous dodging without any human presence in our autonomous vehicle was performed off-roads. In this mode, collision avoidance against aggressive-stochastic-adversarial traffic is the sole responsibility of #AutonomousVehicles. Our autonomous vehicle, Deep Xplorer, was tasked with navigating an off-road trail, avoiding both the static and dynamic obstacles. Furthermore, team members on bikes introduced random cross-traffic interactions by cutting its path at random, creating highly adversarial and stochastic traffic scenarios.

Increasing the stakes even further, Founder & CEO, @sanjeevs_iitr drove our other autonomous vehicle (Xplorer), cutting the path of Deep Xplorer at random -- leaving the task of negotiation, motion planning, and decision making, to ensure collision avoidance and safety of the vehicles and the environment as a sole responsibility of our autonomous driving stack in Deep Xplorer.

This is the first time we performed autonomous dodging without human presence in our autonomous vehicle, pushing the limits of what is possible in the autonomous driving landscape.

#ReinforcementLearning #DeepLearning

[IADN @NewsIADN]

Bhopal-based #SwaayatRobots has successfully demonstrated autonomous driving in off-road conditions on it's Deep Xplorer autonomous vehicle.

[Sanjeev Sharma @sanjeevs_iitr]

Thanks so much @vijayshekhar for sharing our work with the world🙏 This was our 103rd demo on roads in India. We are going to bring much more in the coming weeks :)

[Vijay Shekhar Sharma @vijayshekhar]

Made in India autonomous driving technology by @sanjeevs_iitr !
👏🏼👏🏼👏🏼

[Runtime @RuntimeBRT]

Happy Saturday! Here's the latest from India's tech ecosystem:

1. @swaayatt conducted their first autonomous dodging test with no humans in the car

2. @SarvamAI launched Sarvam Studio for content creators

3. @nothingindia has a line down the street in Indiranagar

[Vijay Shekhar Sharma @vijayshekhar]

And you thought autonomous driving won’t be possible in India ?

Here is classic indian driving in Bhopal by autonomous vehicle.

[Sanjeev Sharma @sanjeevs_iitr]

Autonomous dodging off-roads without any human presence. This is a demonstration of pushing the limits of what is possible in off-road #autonomousdriving.

In this demo, we performed autonomous dodging without any human presence in our autonomous vehicle, that too off-roads. In this mode, collision avoidance against aggressive-stochastic-adversarial traffic is the sole responsibility of #AutonomousVehicles. Our autonomous vehicle, Deep Xplorer, was tasked with navigating an off-road trail, avoiding both the static and dynamic obstacles. Furthermore, team members on bikes introduced random cross-traffic interactions by cutting its path at random, creating highly adversarial and stochastic traffic scenarios.

To increase the stakes even further, I personally drove our other autonomous vehicle (Xplorer), cutting the path of Deep Xplorer at random --leaving the task of negotiation, motion planning, and decision making, to ensure collision avoidance and safety of the vehicles and the environment as a sole responsibility of our autonomous driving stack in Deep Xplorer.

This is the first time in the history of @swaayatt that we performed autonomous dodging without human presence in our autonomous vehicle, pushing the limits of what is possible in autonomous driving landscape.

#reinforcementlearning #deeplearning

[Deep Eigen @DeepEigen]

Off-road L4+ #AutonomousDriving negotiating stochastic, adversarial, bi-directional traffic dynamics — without a safety driver.

A major milestone demonstrated by @swaayatt.

The demo was conducted in two stages. In the first stage, there was no safety driver, with only a passenger present to operate a kill switch as a precaution. The autonomous vehicle navigated off-road trails amid bi-directional traffic, where dynamic obstacles frequently approached from the wrong side, creating adversarial scenarios. The system also performed autonomous off-road dodging, with team members intentionally cutting across its path to introduce unpredictable, real-world interactions.

In the second stage, all human presence was removed from the vehicle. The autonomous agents carried out end-to-end navigation, contingency handling, and collision avoidance entirely onboard.

Off-road L4+ autonomy operates in a fundamentally different regime from HD-map-centric approaches. Such environments demand motion planning and decision-making agents with advanced onboard intelligence, capable of perceiving, reasoning, and negotiating complex, stochastic, adversarial traffic dynamics in real time.

A significant step forward for real-world autonomous systems.

Happy Republic Day 🇮🇳

#AutonomousVehicles #ReinforcementLearning #DeepLearning

[Swaayatt Robots @swaayatt]

Off-road L4+ #AutonomousDriving negotiating stochastic, adversarial, bi-directional traffic dynamics — without a safety driver.

This demo was performed in two parts.

In Part 1, there was no safety driver, but the passenger seat was occupied to operate the kill switch as a precaution. Our autonomous vehicle followed an off-road trail in the presence of bi-directional traffic, where dynamic obstacles often approached from the wrong side, creating adversarial scenarios. The vehicle also performed autonomous off-road dodging. In this mode, collision avoidance is solely the responsibility of the vehicle, while team members intentionally cut across its path at random, creating unpredictable, real-world adversarial interactions.

In Part 2, we removed all human presence from inside the vehicle, leaving it entirely to the autonomous agents to perform end-to-end navigation and handle contingencies independently.

Off-road L4+ autonomy operates in a fundamentally different regime from HD-map-centric approaches. Driving behavior and motion primitives cannot be embedded into structured maps in such environments — HD-map-based navigation and decision-making off-road is not scalable. Motion planning and decision-making agents must instead demonstrate advanced onboard intelligence to perceive, reason, and negotiate complex, stochastic, adversarial traffic interactions in real time.

This demonstration marks the beginning of our large-scale L4+ autonomous driving operations — both on-road and off-road — targeting civilian and military applications.

Happy Republic Day 🇮🇳

#AutonomousVehicles #ReinforcementLearning #DeepLearning

[Sanjeev Sharma @sanjeevs_iitr]

Off-road L4+ #autonomousdriving negotiating stochastic, adversarial, bi-directional traffic dynamics without a safety driver.

For the first time in the history of Swaayatt Robots (स्वायत्त रोबोट्स), we have completely removed the human safety driver from our autonomous vehicle.

This demo was performed in two parts. In the first part, there was no safety driver, but the passenger seat was occupied to press the kill switch in case of an emergency. In the second part, there was no human presence inside the vehicle at all.

In the first part of the demo, our autonomous vehicle was tasked with following an off-road trail in the presence of bi-directional traffic, where dynamic obstacles often approached from the wrong side, creating adversarial scenarios. There was no safety driver in the driver's seat, and only the passenger seat was occupied to press the kill switch in case of an emergency. The vehicle also performed autonomous off-road dodging. In this mode, it is the sole responsibility of the vehicle to ensure collision avoidance, while team members intentionally cut across its path at random, creating unpredictable, real-world adversarial situations.

To increase the stakes further, in the second part we completely removed human presence from inside the vehicle, leaving it entirely to the autonomous agents to perform end-to-end autonomous navigation and handle contingencies on their own.

While many contemporary approaches to L5 autonomy assume the embedding of motion primitives within HD maps, L4+ autonomy off-road operates in a completely different regime. There is no practical provision for embedding driving behavior or motion primitives into an HD map for off-road environments -- HD-map-based navigation and decision-making off-road is not scalable. Motion planning and decision-making agents must instead possess sufficiently advanced onboard intelligence, enabling #autonomousvehicles to perceive, reason, and negotiate complex, stochastic, adversarial traffic interactions in real time.

With this demonstration, we begin our large-scale L4+ autonomous driving operations both on- and off-road, targeting civilian as well as military applications.

I wish everyone a very Happy Republic Day!

#reinforcementlearning #deeplearning

[Sanjeev Sharma @sanjeevs_iitr]

We had Priyanka Chaturvedi with us today at @swaayatt, who is a great friend of mine. We took her for a ride in our autonomous vehicle, demonstrating generic #autonomousdriving, performing obstacles avoidance, and dodging of obstacles.

Autonomous dodging has now become a norm at Swaayatt, where our team members cut the path of our #autonomousvehicles at random, and responsibility for collision avoidance solely rests with the vehicle. At it can be seen in this demo, the vehicle avoided close-calls, where such navigation task would entirely be classified as a corner case in the West.

The core motion planning algorithm that performs this operation runs at 200+ Hz on a single thread of a regular mobile i7 processor. That is an order of magnitude more computationally efficient than anything that exists out there. This directly translates to significantly less energy consumption onboard.

This framework is being converted to deep #reinforcementlearning based framework, solving the integrated planning, controls, and decision making problem.

#deeplearning #adas

[Swaayatt Robots @swaayatt]

With this #Diwali, we begin our #autonomousdriving operations in the regular city traffic, in India.

We will gradually increase the complexity of the traffic in which we validate our technology for large scale deployment globally.

In the first part of the demo, our autonomous vehicle navigated on a regular 2-lane road, negotiating intersections, roundabouts, and regular traffic in its lane.

In the second part, we performed autonomous dodging, where the responsibility of collision avoidance and negotiation rests solely with the autonomous vehicle.

We asked our team members to abruptly cut the path of our vehicle at random, while the regular traffic was also on the road -- they were made aware of the autonomous operations by our team members on the bikes, following our autonomous vehicle.

In the earlier demos, we had controlled traffic scenarios, where during the autonomous dodging by our #AutonomousVehicles, most of the traffic cutting path was generated by our team. This time, it was done in the presence of regular traffic.

To achieve Level-5 autonomy, autonomous vehicles will have to ensure that they are able to respond properly to unforeseeable sudden occurrence of obstacles in its path. The motion planning and decision making capabilities being developed over here, to perform autonomous dodging, will help achieve Level-5 autonomy on Indian roads very soon.

Swaayatt Robots wishes everyone a very Happy and Shubh Diwali! 🙏

Jai Sri Ram 🙏 Jai Maa Kali🙏

#reinforcementlearning #DeepLearning

[Sanjeev Sharma @sanjeevs_iitr]

With this Diwali, we begin our #autonomousdriving operations in the regular city traffic, gradually increasing the complexity of the traffic in which we validate our technology for large scale deployment globally.

In the first part of the demo, our autonomous vehicle navigated on a regular 2-lane road, negotiating three intersections, roundabouts, and regular traffic in its lane.

In the second part, we performed autonomous dodging, where the responsibility of collision avoidance and negotiation rests solely with the autonomous vehicle. We asked our team members to abruptly cut the path of our vehicle at random, while the regular traffic was also on the road.

In the earlier demos, we had controlled traffic scenarios, where during the autonomous dodging by our #AutonomousVehicles, most of the traffic cutting path was generated by our team. This time, it was done in the presence of regular traffic.

To achieve Level-5 autonomy, autonomous vehicles will have to ensure that they are able to respond properly to unforeseeable sudden occurrence of obstacles in its path. This motion planning and decision making capability developed at @swaayatt, to perform autonomous dodging, will help achieve Level-5 autonomy on Indian roads very soon.

I wish everyone in my network a very Happy and Shubh Diwali! 🙏

Jai Sri Ram 🙏 Jai Maa Kali 🙏

#reinforcementlearning #DeepLearning

[Swaayatt Robots @swaayatt]

We were invited at the Military College of Telecommunication Engineering in Mhow, to demonstrate our #autonomousdriving technology to the army officers, accompanied with a technical talk by the founder @sanjeevs_iitr on AI and autonomous navigation, and the cutting-edge research we've been doing on several frontiers to enable #autonomousvehicles robustly negotiate challenging environments, terrains, and traffic scenarios.

A day before the talk and the live demo, we performed geo-fenced navigation, along with negotiation of the generic traffic on open roads in the military college campus.

This involved generic navigation on double lane and single lane roads, generic obstacles avoidance, negotiation of security barriers, gates, and check-posts, and much more -- all of which couldn't be captured in this video, like autonomous dodging which we have shown in the last two demos.

We also discussed how we have been pioneering #ReinforcementLearning for autonomous driving, and tackling autonomous navigation challenges both on- and off-roads.

#deeplearning

[Sanjeev Sharma @sanjeevs_iitr]

We were invited at the Military College of Telecommunication Engineering in Mhow, to demonstrate our #autonomousdriving technology to the senior army officials and infantry officers, along with a technical talk on the artificial intelligence and autonomous navigation, and the research we have been doing on these fronts at @swaayatt.

A day before the trial and live demo we performed geo-fenced navigation, along with negotiation of the generic traffic on open roads in the military college campus. This involved generic navigation on double lane and single lane roads, generic obstacles avoidance, negotiation of security barriers, gates, and check-posts, and much more -- that was not captured in the video, like autonomous dodging which we have shown in the last two demos.

It was an enriching experience to meet the people who protect our borders and ensure that the nation is secure, and to present them with the boundary of knowledge in the respective fields along with the discussions of the research frontiers involved in the development of robust and all terrain #autonomousvehicles. We also discussed how Swaayatt is pioneering #reinforcementlearning for autonomous driving, and tackling autonomous navigation challenges both on- and off-roads.

#deeplearning

@artrac_ia

[Swaayatt Robots @swaayatt]

Human level Control for Autonomous Dodging of Adversarial Traffic with Bi-Directional Negotiation.

In this demo one can see our autonomous vehicle dodging adversarial traffic on a tight single-lane road with muddy side-road, showcasing the robustness of our #autonomousdriving technology.

Earlier, on 27th of August, while demonstrating to representatives from Suzuki Motor Corporation from Japan, we performed low-speed autonomous dodging on a wide road. In this demo, we increased the complexity even further, and last week we performed autonomous dodging and bi-directional negotiation at much higher speeds.

During autonomous navigation, we asked our engineers to use a bicycle, a bike, and an SUV to cut the path of our autonomous vehicle, at random / at will, creating a stochastic-adversarial scenario, where the sole responsibility of collision avoidance and detouring is that of our #AutonomousVehicles .

As a disclaimer,
a) The engineer cutting the path of our autonomous vehicle by overtaking it through a bicycle is the founder @sanjeevs_iitr.
b) The engineer driving the SUV at high-speeds on muddy roads overtaking and then cutting the path of the vehicle is again the Sanjeev Sharma (founder).
c) The engineer who cut the path of our autonomous vehicle at 22-sec, 2:40-mark from Nipun Shah, and the engineer who cut the path with a bike at 1:46-mark is Paras Rana.

Furthermore, our vehicle also performed traffic-cone negotiation, as well as generic bi-directional traffic negotiation on a single lane road with ease -- a capability that we have been a pioneer of and have been demonstrating since October 2023.

The motion planning and decision making framework performing dodging and bidirectional negotiation used one #reinforcementlearning agent as of now. The underlying policy is being scaled up for Level-5 autonomy via deep coordinated multi-agent reinforcement learning.

#deeplearning

[Deep Eigen @DeepEigen]

Human Level Autonomous Driving/Dodging of Adversarial Traffic!

[Sanjeev Sharma @sanjeevs_iitr]

Autonomous Driving: Human level Control for Autonomous Dodging of Adversarial Traffic with Bi-Directional Traffic Negotiation.

In this demo one can see our autonomous vehicle at @swaayatt autonomously dodging adversarial traffic on a tight single-lane road with muddy sides, showcasing the robustness of our #autonomousdriving technology.

Earlier, on 27th of August, while demonstrating to representatives from Suzuki Motor Corporation from Japan, we did a low-speed autonomous dodging on a wide road. We increased the complexity for our autonomous vehicle even further and last week we performed autonomous dodging and bi-directional negotiation at much higher speeds.

While our vehicle was navigating autonomously, we asked our engineering team to use a bicycle, a bike, and an SUV to cut the path of our autonomous vehicle, at random / at will, creating a very challenging adversarial scenario, where the sole responsibility of collision avoidance, detouring is that of our #autonomousvehicles.

As a disclaimer,
a) The person cutting the path of our autonomous vehicle by overtaking it through a bicycle was myself.
b) The person driving the SUV at high-speeds on muddy roads overtaking and then cutting the path of autonomous vehicle is again myself.
c) The team member who cut the path our autonomous vehicle at 22-sec and 2:40-mark in the video is our mechanical engineer Nipun Shah and the team member who cut the path with a bike at 1:46-mark is our robotics engineer Paras Rana.

Furthermore, our vehicle also performed traffic-cone negotiation, as well as generic bi-directional traffic on a single lane road with ease -- a capability that Swaayatt Robots is a pioneer of and has been demonstrating since October 2023.

The motion planning and decision making framework performing dodging and bidirectional negotiation is something I have been personally developing for quite some time now, and it uses one RL agent. This underlying policy is being scaled up for Level-5 autonomy via deep #reinforcementlearning.

#deeplearning

[Sanjeev Sharma @sanjeevs_iitr]

Autonomous Driving: Human level Control for Autonomous Dodging of Adversarial Traffic with Bi-Directional Traffic Negotiation.

In this demo one can see our autonomous vehicle at @swaayatt autonomously dodging adversarial traffic on a tight single-lane road with muddy sides, showcasing the robustness of our #autonomousdriving technology.

Earlier, on 27th of August, while demonstrating to representatives from Suzuki Motor Corporation from Japan, we did a low-speed autonomous dodging on a wide road. We increased the complexity for our autonomous vehicle even further and last week we performed autonomous dodging and bi-directional negotiation at much higher speeds.

While our vehicle was navigating autonomously, we asked our engineering team to use a bicycle, a bike, and an SUV to cut the path of our autonomous vehicle, at random / at will, creating a very challenging adversarial scenario, where the sole responsibility of collision avoidance, detouring is that of our #autonomousvehicles.

As a disclaimer,
a) The person cutting the path of our autonomous vehicle by overtaking it through a bicycle was myself.
b) The person driving the SUV at high-speeds on muddy roads overtaking and then cutting the path of autonomous vehicle is again myself.
c) The team member who cut the path our autonomous vehicle at 22-sec and 2:40-mark in the video is our mechanical engineer Nipun Shah and the team member who cut the path with a bike at 1:46-mark is our robotics engineer Paras Rana.

Furthermore, our vehicle also performed traffic-cone negotiation, as well as generic bi-directional traffic on a single lane road with ease -- a capability that Swaayatt Robots is a pioneer of and has been demonstrating since October 2023.

The motion planning and decision making framework performing dodging and bidirectional negotiation is something I have been personally developing for quite some time now, and it uses one RL agent. This underlying policy is being scaled up for Level-5 autonomy via deep #reinforcementlearning.

#deeplearning