Pulkit Agrawal
@pulkitology
Followers
12K
Following
446
Media
110
Statuses
358
Presenting Visual Dexterity: Object re-orientation in its full generality! Single camera. Novel objects. Any orientation. Downward-facing hand that fights gravity. Real-time dynamic control. Open source setup. Learn more: Led by @taochenshh #robots #rl
5
58
313
DexWrist is built for compliance, force control and is easy to simulate. It uses quasi direct drive actuators, is backdrivable and has high control bandwidth. Work led by @martinpeticco @JohnMarangola and Gabriella. @MIT_CSAIL @csail_alliances @MIT.
1
0
1
DexWrist makes teleop more intuitive, shorter trajectories and thus allows faster and cheaper data collection.
1
0
1
What if robots had wrists? . Introducing DexWrist: a drop-in addition to your robot arm to operate in messy environments, be more dynamic, more teleoperation intuitive and overall better performance. Want to learn more or get a DexWrist?
1
6
48
What if robots had wrists? 👇.
What’s keeping robot arms from working like human arms?. They're big, slow, have the wrong joints, and can't conform to their environment. DexWrist solves all of these issues and simplifies learning constrained, dynamic manipulation👉Â
2
0
42
RT @martinpeticco: What’s keeping robot arms from working like human arms?. They're big, slow, have the wrong joints, and can't conform to….
0
51
0
What if an LLM can decide what data to use, potentially generate its own data and decide how to update itself 👇.
What if an LLM could update its own weights?. Meet SEALđź¦: a framework where LLMs generate their own training data (self-edits) to update their weights in response to new inputs. Self-editing is learned via RL, using the updated model’s downstream performance as reward.
2
3
18
Llama 4 (@Meta) results are consistent with what we hypothesized will unleash the next generation of AI reasoning. A new paradigm for pre-training is around the corner
Llama 4 (@Meta) shows too much SFT limits RL exploration — something we also found in our recent work! A new and superior pretraining paradigm is around the corner to unleash a new era of reasoning. Check out our paper: Thread:
1
3
22
Initial tests, even on pre-trained language models, suggest that directly doing reward-based fine-tuning and skipping supervised fine-tuning works better! . Joint work with @seungwookh, @jyo_pari and @gershbrain.
2
0
18
We tested the reasoning of current large models on simple tasks like sorting and printing numbers, but in new languages not common on the internet -- these models mostly fail!
1
0
21
Overturning the next-token prediction is required for achieving general reasoning! We predict that RPT (Reward Pre-Training) will overtake GPT in the future -- similar to how AlphaZero overtook AlphaGo. Learn more: 🚨Our whitepaper, “General Reasoning
5
60
429
ORSO outperforms other methods of choosing reward functions and vastly outperforms Eureka which uses Naive reward selection mechanisms.
0
0
5
Casting reward selection as a model selection leads up to 8x faster learning and 50% better performance! (.⚡ Provable regret guarantees. 🌟 Easy to implement (. ⚔️ 1 GPU can do the work of up to 8 GPUs! . Presenting ORSO:
3
23
144
Agents performing "curious" exploration by setting random abstract goals! We present a simple but effective method for deep exploration of reinforcement learning (RL) agents that we call random latent exploration (RLE). Typical RL gents explore by:.⚔️ Noise-based exploration,
5
29
186
You aren't doing robotics if you are not breaking some robots!
12
25
370
Presenting Unsupervised Actuator Nets (UANs) that push the limits of agile whole-body control without the need for reward shaping! . ⚡️ UANs reduce the sim2real gap in robot's motors removing the need for reward design to bridge the sim2real gap. ⚡️ A pre-trained whole-body
2
19
152
Auditing and exposing the fragility of language-conditioned robot models with Embodied Red Teaming (ERT)! . 🤯 Simple re-phrasing of task instructions, e.g., from "Please bring me a can of coke" to "Give me a coke," is the difference between the robot succeeding or failing.
1
12
78