Sumedh Reddy

Research at TDK America R&D | previously at UMD Robotics, Open Robotics

[Sumedh Reddy @RKSumedh]

@ChongZzZhang @JayHe748646 @KehanWen170077 @leggedrobotics 10+ ladder climbs in a row on 4cm rungs is wild. how few human demos are we talking for the warmstart, and does the contact planner ever find sequences that weren't in them?

[Sumedh Reddy @RKSumedh]

@Jiahui_Yang6709 splitting it into wbc for the reach and rl for the contact part makes a lot of sense. after distillation does the policy ever blend the two, like start shaping the grasp before the reach is done, or is the handoff still pretty visible?

[Sumedh Reddy @RKSumedh]

@yeokch yeah that tradeoff is the scary part. smooth enough to get useful gradients and the policy might lean on contacts that don't really exist on hardware. would love to see a getup clip either way

[Sumedh Reddy @RKSumedh]

finetuned as a policy it's competitive with Cosmos Policy on LIBERO and RoboCasa, at up to 5.5x faster inference.

my take: text as a shared action space is the real idea. it lets unlabeled video into world model training. caveat: policy results are sim only.

[Sumedh Reddy @RKSumedh]

naver ai lab's LaLa-World writes robot actions as plain text ("move the pusher 1.19 left and 0.68 up") so one JEPA-style latent world model can train across 7 control datasets plus regular video. the predictor is Qwen3.5-0.8B on frozen DINO features.

https://t.co/8PjpHmLSFF

[Sumedh Reddy @RKSumedh]

no pixel reconstruction, just next-latent prediction. planning: 72.1% avg on 5 sim tasks, 7.5 pts over JEPA-WMs, push-t 68.8% to 81.2%. against RAEv2, a pixel-generative model of similar size on the same data: better planning, 6x cheaper training, 13.9x faster.

[Sumedh Reddy @RKSumedh]

@mkirchhof_ kind of surprised the harsh step penalties win, i'd have guessed smoother ones train more stably. with gate does the model learn to actually plan inside the budget, or does it mostly just give up early on hard tasks?

[Sumedh Reddy @RKSumedh]

@realyunlong the realistic forward, simplified differentiable backward split is neat. how far apart can the two physics get before the gradients start pointing the wrong way, especially around contacts?

[Sumedh Reddy @RKSumedh]

my take: per-clip advantage centering is a cheap trick more multi-task RL should copy. caveats: the numbers are all sim, one rigid object, baselines adapted by the authors. their own 1,575-clip dataset only adds ~2 points. real two-hand transfer is the test that matters.

[Sumedh Reddy @RKSumedh]

ziwei liu and the ropedia team trained one policy that makes two dexterous robot hands follow human two-hand manipulation clips, including motions and objects it never trained on. one checkpoint, no finetuning. most dexterous trackers are single hand.

https://t.co/faB3PeTCGu

[Sumedh Reddy @RKSumedh]

the fix i liked most: subtract each clip's mean advantage, so hard clips don't get uniformly negative updates. dropping it takes unseen objects from 57.3% to 48.2%. baselines (two-hand versions of dexplore and dextrack) get 11.6% and 0%. on unseen trajectories, 86.8% vs 9.4%.

[Sumedh Reddy @RKSumedh]

@mingqi70158 curious if the savings hold up when the judge's bias isn't uniform across systems, like when it quietly favors outputs from its own model family. does PPI still correct for that?

[Sumedh Reddy @RKSumedh]

@chelseabfinn @Anikait_Singh_ @teresaszhang @yoonholeee @roshen_nair @aviral_kumar2 curious how this holds up when the corpus keeps changing. do you have to re-optimize the organization every time new docs come in or can it update incrementally?

[Sumedh Reddy @RKSumedh]

@GuanyaShi @CMU_Robotics this is awesome. are the labs sim first and then onto the real robot, or straight to hardware? curious how you keep that many small humanoids alive for a whole semester

[Sumedh Reddy @RKSumedh]

@make_robot_go @roboticscorp the putting things back where they belong part seems like the hard bit to me. how does voxel figure out where a thing lives in a home it's never seen?

[Sumedh Reddy @RKSumedh]

@jqdsouza how does beacon decide what's worth remembering from a session? i'd worry about it holding onto a bad workaround it picked up once and then repeating it

[Sumedh Reddy @RKSumedh]

flapping airplanes raised a $1B series A, co-led by index and lux, to build new gpu primitives that change how machines learn. huge number for an ai lab. curious if the bet is mostly about data efficiency https://t.co/ttGq70F8Fp

[Flapping Airplanes @flappyairplanes]

We’ve spent the past year developing new primitives for GPUs. We believe this enables new algorithms to change how machines learn, open new markets, and dramatically reshape AI’s cost structure.

We have raised a $1 billion Series A, co-led by Index and Lux.

[Sumedh Reddy @RKSumedh]

@pabbeel does the policy ever start exploiting the world model once you backprop through it? like finding moves where the gradients look great but the real sim disagrees. curious how often you have to refit the model