Robot Demos

Real-robot demos — UniWALM manipulation (bimanual, dexterous, memory-dependent, RMBench) and self-driving car work.

Short demo reels from recent and past work — UniWALM (real-robot manipulation) and a self-driving car project. For full write-ups: UniWALM project page, Self-Driving project page.

Note on UniWALM videos below. Real-robot rollouts here are from an unoptimized deployment pipeline (I/O, teleop stack, logging overhead) and run without a real-time controller (RTC), which is why the motion appears jerky. The actual UniWALM policy + language-memory reasoning runs at ~140 ms per step (N = 10 flow-matching steps on an H100). Smoothness would come for free with an RTC layer on top.

Capability wall

Overview reel across the four UniWALM task families: bimanual, dexterous, multi-stage, and memory-dependent.

Capability wall across tasks. 3× speed.

RMBench — memory-dependent benchmark

UniWALM reaches 34.5% on RMBench's memory-dependent M(N) subset — +21.2 points over same-backbone FastWAM (13.3%), and exceeds Mem-0's privileged-keyframe setup (28.5%) without any pretrained VLM. Below: 12 real-robot successes sampled across the four tasks, randomly ordered.

12 successful real-robot rollouts across Press Button, Battery Try, Blocks Ranking Try, and Cover Blocks, randomly ordered in a 4×3 grid. 6× speed.

Multi-stage dexterous bimanual — item removal

Sequential bimanual removal of 5 items from a container across varied object layouts. Deployed on Flexiv bimanual arms with Sharpa / Paxini hands.

Multi-stage item removal rollouts. 3× speed.

Find-the-Item — memory-dependent manipulation

A real-world Find-the-Item task designed to require memory: the policy sees the target object once at episode start, then has to recover it later across visual distractors. Generating a textual memory lifts success from 11.1% (FastWAM, same video backbone, no language) to 61.1% — state-of-the-art VLAs stay at or below chance (≤ 33.3%).

Find-the-Item rollouts. The policy re-reads its own natural-language memory on every control step. 3× speed.

Self-driving car — RRC IIIT (2020–2021)

Earlier work from undergrad / M.Sc. at the Robotics Research Center, under Prof. K. Madhava Krishna. I designed the mapping and obstacle-avoidance modules and built the ROS pipeline that fused perception outputs into the occupancy map used by the planner.

Autonomous driving (open field). Closed-loop mapping + planner driving the vehicle. 2× speed.
Obstacle avoidance. Local-planner reaction to detected obstacles. 2× speed.
Drive-by-wire via Xbox controller. A fun early milestone — wiring the car up end-to-end so it could be driven over a joystick before autonomy was layered on top. Became the teleop + safety-override channel for later testing.
Full project pages: UniWALM → · Self-Driving →