
BUILDING INDIA’S AI ENGINE
Reliable Robots, Deployed in the Real World
Most people encounter AI through a chat interface.
The harder challenge is getting intelligence into a
body that moves through the physical world; it may
be less visible but is no less consequential. At
Plaksha's AI and Robotics Lab, that is exactly the
problem under investigation.
The lab's research is organized around a key
objective – reliability of robots. Three questions
drive it: how do you systematically teach robots
versatile motor skills? How do you get them to plan
reliably over a longer horizon in complex
environments? And how can they understand and
reason about the world around them? Versatile
motor skills, long-horizon planning, and embodied
reasoning define the lab's agenda.
Legged robots are the test case of choice. They are
harder to control but are far more versatile,
capable of climbing stairs, crossing rubble, and
adapting to surfaces. However, the dominant
training method, imitation learning, has a
weakness: models trained by copying expert
demonstrations do not generalize well. They are
brittle in novel scenarios and poorly grounded in
physics. The lab's research targets these gaps.
On the locomotion side, we have developed and
deployed a quadruped locomotion policy trained
using curriculum-based reinforcement learning in
Isaac-Sim. The work tests multiple architectures,
blind policies relying only on proprioception,
terrain-aware policies with explicit geometry
sensing and adaptive variants that adjust to mid-
episode friction changes, establishing clear
performance tradeoffs between them. Physical
deployment on a quadruped robot has been
achieved, with a structured sit-to-stand transition
added to the pipeline. Rahath Malladi, a fourth-
year BTech student, extends this work toward
multi-embodiment generalization: a single policy
trained to walk Anymal, Spot, and the H1
humanoid across a shared simulation
environment.
The lab also works on the planning layer. Bharat
Jain, a third-year BTech student, has built a VLM-
based autonomous navigation system that
connects Claude to ROS via MCP, enabling a robot
to explore an indoor environment, build occupancy
map, label it using a vision-language model, and
navigate to any object a user names in natural
language. This framework has been shown to work
in the Gazebo simulation environment. Separately,
Arunachalam Vijayanand, a second-year BTech
student, is fine-tuning LLMs for direct robot control
using reinforcement learning techniques. The goal
here is to develop a framework to tune LLM or VLM
as that would enable them to be grounded better
in the physics that underpins the robot and its
interactions with the real world. In some initial
experiments, the lab was to train an LLM to reach a
52% goal-completion rate on a drone maze task,
matching state-of-the-art performance.
Cutting across all three streams is a model-based
learning approach being developed in collaboration
with UC San Diego, and UC Irvine. World models by
their very nature are built to replicate the physics
of the real world and they in principle can be
learned by an agent itself by self-exploration. Such
a world-model, in principle, removes the grounding
problem and affords the agent the ability to reason
and plan by simulating the world model. The lab
has shown that a world model learned entirely
through self-exploration without any prior data can
support policy learning that is significantly more
efficient compared to existing state-of-the-art
model-based learning approaches and holds up
under out-of-distribution perturbations.
The work is incremental in the best sense: each
project addresses a specific, well-defined gap.
Together, they are laying groundwork for robots
that can be deployed in the real world.
Deepan Muthirayan
Assistant Professor
Plaksha University
10
2014
I Learned to Imagine
Two landmark deep generative models arrived.
Generative Adversarial Networks (GANs), introduced by
Ian Goodfellow and colleagues, pitted a generator
against a discriminator in an adversarial game, each
improving the other until the generator could produce
images indistinguishable from reality. Variational
Autoencoders (VAEs) offered a complementary
approach: learning compact latent representations of
data and sampling from them to generate diverse new
content.
However, GANs were difficult to train and prone to
mode collapse. VAEs produced diverse results, often
with unwanted blurriness.
C R E A T E
Building India’s AI Engine | Alignment is the Key to Safe AI | AI for Humanity | AI as a Tool to Question | AI Ecosystem - People, Partnerships and Platforms | AI Infrastructure |
Developing Algorithms for Economic Value | Student Innovations
BTech students at work in the ARK Robotics Lab