In my latest PhD paper, we declare WAR on sensor-maxxing. For the...
Aditya Bhatt ✈️ IROS 2026@aditya_bhatt
69 views
Sep 26, 2026
~2 min read
Advertisement
1
In my latest PhD paper, we declare WAR on sensor-maxxing.
For the first time, push-resilient humanoid walking, just with joint encoders! No IMU, no F/T sensors.
🥁 Introducing Blind Dexterity 🧵 👇
w/ @OKaidanov @liu_puze @Jan_R_Peters at @DFKI @ias_tudarmstadt
For the first time, push-resilient humanoid walking, just with joint encoders! No IMU, no F/T sensors.
🥁 Introducing Blind Dexterity 🧵 👇
w/ @OKaidanov @liu_puze @Jan_R_Peters at @DFKI @ias_tudarmstadt
0:24
2
2/ 😇 Sensors are good, actually. Please don’t dismantle your robot! 😅
We impose sensory deprivation and delete all exteroception—pose, vision, tactile sensors, force/torque—and ask:
How much dexterous manipulation can RL squeeze out of proprioception alone?
A lot. 🦾
We impose sensory deprivation and delete all exteroception—pose, vision, tactile sensors, force/torque—and ask:
How much dexterous manipulation can RL squeeze out of proprioception alone?
A lot. 🦾
3
3/ Task 1: Trap a football ⚽, somewhere in front of the robot, with one foot.
Emergent behavior: The policy learns to search: sweep and wiggle the foot, make contact, and trap it.
Deprived of direct ball position sensing, it must earn that info by poking the world.
Emergent behavior: The policy learns to search: sweep and wiggle the foot, make contact, and trap it.
Deprived of direct ball position sensing, it must earn that info by poking the world.
0:11
4
4/ Task 2: Blindly board a thing with wheels. 🛹 😉
The robot touches and perturbs the skateboard to figure out its position and orientation, then hops on.
Searching for your support surface while actively moving it around. Somehow, this works. 😅
The robot touches and perturbs the skateboard to figure out its position and orientation, then hops on.
Searching for your support surface while actively moving it around. Somehow, this works. 😅
0:08
5
5/ Task 3: Find and lift a suitcase 🧳 handle by feel
Random table heights, random placement on the table.
"feel" = joint angle vibes
Random table heights, random placement on the table.
"feel" = joint angle vibes
0:15
6
6/ Why does any of this work?
We argue that with compliant PD control, joint-angle deviations during contact carry clues about the world.
Joint angles are a sparse "haptic" channel. Poking and smacking is a perception strategy, and returns useful proprioceptive signatures.
We argue that with compliant PD control, joint-angle deviations during contact carry clues about the world.
Joint angles are a sparse "haptic" channel. Poking and smacking is a perception strategy, and returns useful proprioceptive signatures.
7
7/ How useful are those proprio signatures? We train state estimators to decode object pose from them.
The green render is the max-likelihood estimate given the last 0.1s of proprioception.
The behavior is surprisingly sensible. Predictions update meaningfully with contacts!
The green render is the max-likelihood estimate given the last 0.1s of proprioception.
The behavior is surprisingly sensible. Predictions update meaningfully with contacts!
8
8/ Sensible Interactive Perception Behavior
Exhibit: Suitcase lifting (randomized table heights, randomized suitcase poses).
The robot first smacks one hand on the table, which clarifies the table height, and instantly localises the ghost suitcase to be flush on top of it.
Exhibit: Suitcase lifting (randomized table heights, randomized suitcase poses).
The robot first smacks one hand on the table, which clarifies the table height, and instantly localises the ghost suitcase to be flush on top of it.
0:33
9
8/ Important Takeaway: Blind RL from scratch worked way better than imitating a pose-aware teacher
The teacher knows exactly where to reach, which isn't distillable into a partially-observing student!
Active perception did not emerge from distillation.
The fumbling is useful!
The teacher knows exactly where to reach, which isn't distillable into a partially-observing student!
Active perception did not emerge from distillation.
The fumbling is useful!
10
9/ There’s a lot hiding in your joint encoders—even before you add richer sensing.
Paper: aditya.bhatts.org/BlindDexterity/
arxiv: arxiv.org/abs/2608.29487
I’ll be at #IROS2026 workshops. Into humanoids? Dexterous manipulation? Slightly weird control strategies? Let's chat.
Paper: aditya.bhatts.org/BlindDexterity/
arxiv: arxiv.org/abs/2608.29487
I’ll be at #IROS2026 workshops. Into humanoids? Dexterous manipulation? Slightly weird control strategies? Let's chat.