Tirth Gada (@tirth_gada) profile picture

Speedrunning Robotics intern @ Human-interactive Robotics Lab, IISc robot learning @SRAVJTI

Mumbai
Joined August 2024
Photo by Tirth Gada (@tirth_gada)Photo by Tirth Gada (@tirth_gada)Photo by Tirth Gada (@tirth_gada)Photo by Tirth Gada (@tirth_gada)Photo by Tirth Gada (@tirth_gada)Photo by Tirth Gada (@tirth_gada)Photo by Tirth Gada (@tirth_gada)Photo by Tirth Gada (@tirth_gada)Photo by Tirth Gada (@tirth_gada)Photo by Tirth Gada (@tirth_gada)
We all have seen flow policies perform amazingly in generative modeling. But what about pure robotics? Introducing Flow-Skill: FPO++ reinforcement learning for robot manipulation built on ManiSkill3.
3
16
3
107
12,492
Tirth Gada retweeted
Today, Oru'el announces its first Tier-1 SOTA. @oruel_inc just beat @nvidia and @OpenAI on RoboRewardBench. Kanook by Oru'el automates the annotation required to watch every episode to decide if the robot actually did the task or not. SOTA: Kanook by Oru'el 30% better than Nvidia FLORA 50% better than OpenAI GPT-6 Astra
4
5
1
33
1,359
Tirth Gada retweeted
Drones are hard to build and control with their hardware debugging weight constraints and shipment delays. But once you cross that the rewards are so very sweet. Here's a blog about my journey with drones from being confused to being confused with obsession.
16
7
61
1,424
the end game is robust tactile sensors to mount on every end effector ever
4
33
910
was having way too much fun with dream1
15
335
autonomous drones are so much cooler than just arms – at Mumbai, India
figure eight trajectory through RAPTOR policy. (a lil rough around edges)
1
13
373
on a mission to make robotics a little less mysterious. one build at a time :) – at Mumbai, India
introducing yumi a hacky $50* umi for the YAMs. collect data quickly -> visualize -> train & deploy *camera's not included - used from the other yam lol
2
29
1,551
Did robot learning from first principles as my starter project with @SRAVJTI last year no better feeling than seeing my mentees cooking so hard @aryankakad57 @mahi_zade0712
I always wondered what would have happened if the robots also had the sense of touch. So with that curiosity in mind, me and @mahi_zade0712 worked upon the "mission mimosa". result? policies converge faster and are sample efficient due to the flexi-tac on SO101.
2
2
32
1,349
the best person to teach you robot learning
wrote a little blog summarizing all the robotics and intelligent systems projects i have done in past 2 years with my friends. this summarizes all the things that worked, lessons i have learned and what i think is worth pursuing ahead. link below
12
468
here is to progressing robot learning with @SRAVJTI – at Mumbai, India
we just oss'ed the robot control stack for achieving this - Meta Quest VR teleoperation - Single & dual-arm YAM/SO101 control - first class MolmoAct2 policy inference - Human-in-the-loop + DAgger data collection - data & rollout recording in @LeRobotHF v3 - based on the openpi-control interface from @physical_int github.com/sra-vjti/karma
11
543
this making me fall in love with robotics even more – at Mumbai, India
cloth folding robot. with @tirth_gada
1
1
40
3,610
lessgoo – at Mumbai, India
tshirt folding, pi0.5
14
449
took more time but finally got flexitac working visuotacrile ACT deployed at like half the steps required to converge from normal – at Mumbai, India
Flexitac loaded tomorrow's to do : run visuotactile ACT
3
2
31
1,557
@binghao_huang really fun experience setting this up We have an issue that we are able to just render the tactile map in 12 *16 and not complete 12*32 Asked on discord too hopefully we will be able to debug it sooner – at Mumbai, India
1
3
231
The amount of time I have spent staring at this set up since last week is insane Really funn – at Mumbai, India
zero shot towel folding attempt. RL next
31
2,865
Flexitac loaded tomorrow's to do : run visuotactile ACT – at Mumbai, India
1
20
2,201
trained a Dreamer-4 world model on 240 ABC-130k episodes across 48 tasks, streamed from the 33.8 TB rerun bucket. On one RTX 4090, doubling the tokenizer from 96 → 192 latents improved reconstruction from 22.90 → 28.31 dB.
6
5
1
90
10,483
But better reconstruction didn't fix action conditioning. True actions produced 3.12× ground-truth motion, while shuffled actions produced 2.34×. The model can reconstruct what it sees, but isn't reliably using actions to predict what happens next.
1
11
550
abc's dataset in rerun is so easy to access will have much better results tomorrow with dreamer 4
this is so good at just 3000 steps any better and you would really need to spend time to spot which one is the ground truth
2
3
23
2,621
is it like some crazy skill issue or teleoperating in sim is actually tougher than irl
4
22
2,523